Difference between revisions of "VPP/AArch64"

From fd.io
< VPP
Jump to: navigation, search
(→‎Meeting Minutes)
(→‎Meeting Minutes)
Line 484: Line 484:
 
**** Performance number: NEON>SVE-VLA>SVE-VLS on FPGA without DMC-620
 
**** Performance number: NEON>SVE-VLA>SVE-VLS on FPGA without DMC-620
 
**** FPGA team promises to provide FPGA image with DMC-620
 
**** FPGA team promises to provide FPGA image with DMC-620
−
**** Liangxing will help to benchmark VPP with FPGA enbaled with DMC-620
+
**** Liangxing will help to benchmark VPP with FPGA enabled with DMC-620
−
** VPP IPv4 fragmentation
+
** VPP IPv4 fragmentation - Tianyu & Jieqiang
−
** VPP IPv6 Benchmarking and Profiling
+
*** Add multi-arch support for ip4-frag node but see no perf bump
 +
*** Apply loop unrolling on ip4-frag node
 +
** VPP IPv6 Benchmarking and Profiling - Jieqiang
 
*** IPv6 profiling
 
*** IPv6 profiling
−
**** Hotspot function - ip6_lookup_node/ip6_rewrite_node
+
**** No perf bump for lookup_x2 function in Fd.io gerrit
−
**** Performance degradation with quad loop unrolling applied on ip6_lookup_inline
+
−
**** Patch the current kernel to enable perfmon plugin on VPP
+
−
**** Need to check performance for IPv6 subnet routing
+
 
**** Try Mellaonx nics for IPv6 routing tests
 
**** Try Mellaonx nics for IPv6 routing tests
−
+
** CNF PoC proposal preparation- Tianyu
−
** VPP memif - Tianyu
+
*** Add support for VPP aarch64 docker image build
−
*** CNF PoC proposal preparation  
+
*** Calico use cases exploration on VPP
−
***** Add support for VPP aarch64 docker image build
+
*** Improve ansible scripts to deploy VPP&snort on K8S pods automatically
−
***** Calico use cases exploration on VPP
+
−
***** Improve ansible scripts to deploy VPP&snort on K8S pods automatically
+
 
** VPP IPsec on Arm - Govind
 
** VPP IPsec on Arm - Govind
 
*** Work on IPsec input - Zach
 
*** Work on IPsec input - Zach
Line 508: Line 505:
 
**** Modify the commit message and upstream the perfmon patch - Zach
 
**** Modify the commit message and upstream the perfmon patch - Zach
 
**** Building Intel QAT driver on arm to test IPsec crypt - Picchi
 
**** Building Intel QAT driver on arm to test IPsec crypt - Picchi
−
 
 
'''11/23/2021'''
 
'''11/23/2021'''
 
* Attendees
 
* Attendees

Revision as of 07:37, 1 December 2021

Get Involved

Meeting Details

IRC Channel

#fdio-arm on freenode.net

Slack

Request invitation at https://slack.fd.io/

Jira

Jira issues with ARM64 label

Presentations

Release Milestones

18.10

18.07

18.04

  • CI
    • Upstream patch verification on ARMv8 machines
    • .deb packages

Machines

The FD.io lab is hosted at VEXXHOST colocation centre in Montreal Québec, Canada.

Platform Role Status Hostname IP IPMI Cores RAM Ethernet Distro
Marvell ThunderX VPP dev debug server Running vpp-marvell-dev 10.30.51.38 10.30.50.38 96 128GB 3x40GbE QSFP+ / 4x10GbE SFP+ Ubuntu 18.04.4
CI build server Running in Nomad s53-nomad 10.30.51.39 10.30.50.39 96 128GB 3x40GbE QSFP+ / 4x10GbE SFP+ Ubuntu 18.04.4
CI build server Running in Nomad s54-nomad 10.30.51.40 10.30.50.40 96 128GB 3x40GbE QSFP+ / 4x10GbE SFP+ Ubuntu 18.04.4
CI build server Running in Nomad s52-nomad 10.30.51.65 10.30.50.65 96 256GB 2xQSFP+ / USB Ethernet Ubuntu 18.04.4
CI build server Running in Nomad s51-nomad 10.30.51.66 10.30.50.66 96 256GB 2xQSFP+ / USB Ethernet Ubuntu 18.04.4
CI build server Running in Nomad s49-nomad 10.30.51.67 10.30.50.67 96 256GB 2xQSFP+ / USB Ethernet Ubuntu 18.04.4
CI build server Running in Nomad s50-nomad 10.30.51.68 10.30.50.68 96 256GB 2xQSFP+ / USB Ethernet Ubuntu 18.04.4
Marvell ThunderX2 Perf DUT candidate Running s27-t13-sut1 10.30.51.69 10.30.50.69 224 128GB 3x40GbE QSFP+ XL710-QDA2 Ubuntu 18.04.2
VPP device server Running in Nomad s55-t36-sut1 10.30.51.70 10.30.50.70 256 256GB 2x40GbE QSFP+ XL710-QDA2 / 2x10/25GE SFP+ ConnectX5 Ubuntu 18.04.4
VPP device server Running in Nomad s56-t37-sut1 10.30.51.71 10.30.50.71 256 256GB 2x40GbE QSFP+ XL710-QDA2 / 2x10/25GE SFP+ ConnectX5 Ubuntu 18.04.4
Huawei TaiShan 2280 CSIT testbed Running in CI s17-t33-sut1 10.30.51.36 10.30.50.36 64 128GB 2x10GbE SFP+ Intel X520-DA2 / 2x25GbE SFP28 Mellanox CX-4 18.04.1
CSIT testbed Running in CI s18-t33-sut2 10.30.51.37 10.30.50.37 64 128GB 2x10GbE SFP+ Intel X520-DA2 / 2x25GbE SFP28 Mellanox CX-4 18.04.1
Marvell MACCHIATObin N/A Decommissioned s20-t34-sut1 10.30.51.41 10.30.51.49, then connect to /dev/ttyUSB0 4 16GB 2x10GbE SFP+ Ubuntu 16.04.4
N/A Decommissioned s21-t34-sut2 10.30.51.42 10.30.51.49, then connect to /dev/ttyUSB1 4 16GB 2x10GbE SFP+ Ubuntu 16.04.5
N/A Decommissioned fdio-mcbin3 10.30.51.43 10.30.51.49, then connect to /dev/ttyUSB2 4 16GB 2x10GbE SFP+ Ubuntu 16.04.5
Power Cycler Operational 10.30.50.80
SoftIron OverDrive 1000 N/A Decommissioned softiron-1 10.30.51.12 N/A 4 8GB openSUSE
N/A Decommissioned softiron-2 10.30.51.13 N/A 4 8GB openSUSE
N/A Decommissioned softiron-3 10.30.51.14 N/A 4 8GB openSUSE

Note: to get lab access, create a gpg key, upload it to keyserver, have it signed by a trusted anchor in a video call (fingerprint will be needed) and then an ARM authority (Tina) needs to send an e-mail to helpdesk@fd.io with your name, e-mail, keygrip and key fingerprint

CI

Covers automated build, unit test, and packaging for various Linux distros on ARMv8 machines.

Jenkins job Status Description
vpp-arm-verify-master-ubuntu1604 Running xxx
vpp-arm-merge-master-ubuntu1604 Running xxx
vpp-arm-verify-1804-ubuntu1604 Running xxx
vpp-arm-merge-1804-ubuntu1604 Running xxx

Next steps:

  • make test added to verify jobs
  • Clang build
  • openSUSE Leap 15 | CentOS 7 | Ubuntu 18.04
  • vpp-csit-verify-virl-master or equivalent CSIT functional testing

CSIT

Covers automated functional and performance integration testing on ARMv8 3-node and 2-node testbeds.

https://wiki.fd.io/view/CSIT/AArch64

Contiv-VPP

This Kubernetes network plugin uses FD.io VPP to provide network connectivity between PODs.

https://github.com/contiv/vpp

The installation guide of Contiv-VPP on Arm64 platform is

https://github.com/contiv/vpp/blob/master/docs/arm64/MANUAL_INSTALL_ARM64.md

Porting and Tuning Roadmap

  • VPP Vectorization: Expanding the Neon Library for IPv4 forwarding code path - Sirshak/Lijian
  • Tuning the quad loop/dual loop for small cores - Lijian
  • General performance analysis and tuning of various graph nodes for IPv4 forwarding test case - Sirshak/Lijian
  • Memory Ordering - Sirshak
  • CSIT Performance Test - Khemendra
  • CSIT Device Test - Juraj
  • CSIT Path Test - Juraj

Known Issues

GCC 5.3 ICEs during FP register allocation. Please use GCC 5.4 or newer.

Activity

Recent Patches

snort: fix unused result warning for gcc-10 Merged 11/06 Tianyu Li
l2: fix array-bounds error for prefetch on Arm Merged 11/07 Tianyu Li
ip6: fix IPv6 address calculation error using "ip route add" CLI Merged 10/21 Jieqiang Wang
ipsec: Performance improvement of ipsec4_output_node using flow cache Merged 10/13 Govindarajan Mohandoss
build: fix centos rpm build Merged 10/08 Tianyu Li
vppinfra: fix potential memory access error in _pool_init_fixed Merged 10/05 Jieqiang Wang
svm: fix asan check failed @svm_map_region on arm Merged 06/24 Tianyu Li
l2: fix vrrp prefix mac comparison Merged 06/09 Tianyu Li
build: fix build error after make wipe Merged 06/04 Tianyu Li
memif: fix input node buffer prefetch Merged 05/21 Tianyu Li
memif: fix gcc-10 build error on arm platform Merged 05/21 Tianyu Li
papi: fix ubuntu 1804 make test socket.close error Merged 04/16 Tianyu Li
rdma: fix skip_ipv4_cksum behavior in scalar path Merged 04/15 Tianyu Li
vppinfra: correct intrinsic called by u16x16_from_u8x16 Merged 04/15 Lijian Zhang
vppinfra: fix compiling error due to incompatible udphdr field names Merged 03/05 Jieqiang Wang
avf: optimized with NEON SIMD instruction Merged 12/18 Lijian Zhang
ip: fix compiling error with gcc-10 Merged 09/01 Jieqiang Wang
build: Fix 'make install-deps' errors on aarch64 CentOS 7 Merged 07/29 Jieqiang Wang
acl: correct acl vat help message Merged 07/24 Lijian Zhang
build: add libssl-dev library for ubuntu 20.04 Merged 06/04 Jieqiang Wang
dpdk: fix compiling issue with clang Merged 05/08 Lijian Zhang
vppinfra: fix u32x4_byte_swap on Arm Merged 05/08 Lijian Zhang
build: support arch-specific compiling for Neoverse N1 Merged 04/30 Lijian Zhang
dpdk: false link down issue with ixgbe NIC Merged 03/23 Lijian Zhang
vlib: fix error when creating avf interface on SMP system Merged 03/21 Jieqiang Wang
vlib: leave SIGPROF signal with its default handler Merged 03/21 Jieqiang Wang
build: add libssl-dev for ubuntu 16.04 and 18.04 Merged 03/11 Jieqiang Wang
vlib: fix code of getting numa node with specific cpu_id Merged 02/17 Lijian Zhang
docs: add physmem section in configuration parameters Merged 12/19 Jieqiang Wang
vlib: add max-size configuration parameter for pmalloc Merged 12/18 Jieqiang Wang
crypto: not use vec api with opt_data[VNET_CRYPTO_N_OP_IDS] Merged 11/13 Lijian Zhang
acl: add missing square brackets to vat_help option in acl api Merged 10/31 Jieqiang Wang
dpdk: apply dual loop unrolling in DPDK TX Merged 09/12 Lijian Zhang
ip: apply dual loop unrolling in ip4_rewrite Merged 09/12 Lijian Zhang
ip: apply dual loop unrolling in ip4_input Merged 09/12 Lijian Zhang
build: fix running error with vmxnet3_test_plugin.so Merged 09/11 Jianlin Lv
build: fix unsupported CMake comparison operation Merged 09/05 Jianlin Lv
tap: fix tap interface not working on Arm issue Merged 09/04 Lijian Zhang
build: fix vpp compilation failure on ThunderX2 and Amp Merged 08/19 Jianlin Lv
vppinfra: Update "show cpu" output for AArch64 chips Merged 08/19 Nitin Saxena
vppinfra: refactor test_and_set spinlocks to use clib_spinlock_t Merged 08/02 Jason Zhang
vppinfra: added performance test for clib_rwlock_t (test_rwlock.c) Merged 08/02 Jason Zhang
vppinfra: refactor clib_rwlock_t to use single condition variable Merged 08/02 Jason Zhang
vppinfra: refactor clib_spinlock_t to use compare and swap Merged 08/02 Jason Zhang
vppinfra: added lock performance test for clib_spinlock_t (test_spinlock.c) Merged 08/02 Jason Zhang
vppinfra: refactor use of CLIB_MEMORY_BARRIER () Merged 08/02 Jason Zhang
vppinfra: conformed spinlocks to use CLIB_PAUSE Merged 08/02 Jason Zhang
vppinfra: add u64x2_scatter/u32x4_scatter Merged 06/21 Lijian Zhang
vppinfra: add u64x2_gather/u32x4_gather Merged 06/21 Lijian Zhang
fix compiling error with marvell pp2 plugin Merged 06/11 Jianlin Lv
Switch atomic release API from __sync to __atomic builtin Merged 06/05 Sirshak Das
Switch atomic test and set API from __sync to __atomic builtin Merged 06/05 Sirshak Das
Build packages for generic Arm architecture Merged 05/15 Lijian Zhang
Enable NEON instructions in memcpy_le Merged 05/01 Lijian Zhang
svm_fifo rework to avoid contention on cursize Merged 04/17 Sirshak Das
Re-enable aarch64 neon instruction in vlib_buffer_free_inline Merged 03/20 Lijian Zhang
sctp chunk_len fix Merged 03/06 Sirshak Das
Use acquire/release ordering when accessing svm_fifo shared variable cursize Merged 11/29 Sirshak Das
Optimize xxx_zero_byte_mask NEON function. Merged 11/07 Lijian Zhang
Enable atomic swap and store macro with acquire and release ordering. Merged 11/03 Sirshak Das
Add and enable msb mask vector intrinsic for aarch64. Merged 10/31 Lijian Zhang
vppinfra: add atomic macros for __sync builtins Merged 10/19 Sirshak Das
vppinfra: Fix extendto_high aarch64 NEON api. Merged 10/09 Sirshak Das
Support dynamic dual/quad loop selection on aarch64 Merged 10/01 Lijian Zhang
Enable verbose output during VPP cmake compiling Merged 9/25 Lijian Zhang
dpdk_plugin: fix mlx5 build and runtime issues Merged 9/27 Sirshak Das
Add and enable u32x4_extend_to_u64x2_high for aarch64 NEON intrinsics. Merged 9/12 Sirshak Das
Add horizontal add (hadd) vector intrinsic via NEON. Merged 9/11 Sirshak Das
Add u32x4_extend_to_u64x2 for aarch64 using NEON intrinsics Merged 9/11 Sirshak Das
Replacing vtbl NEON intrinsic with rev NEON intrinsic for byte_swap. Merged 9/11 Sirshak Das
Fix array bound failure in api_sr_localsid_add_del Merged 8/30 Lijian Zhang
cmake: fix marvell plugin build Merged 8/28 Brian Brooks
fix dpdk_plugin.so load failure with DPDK 18.08 Merged 8/23 Lijian Zhang
Fix a bug in function pipe_rx Merged 8/17 Lijian Zhang
fix compiling warnings with GCC Merged 8/17 Lijian Zhang
Update AArch64 CSIT machines into FD.io VPP docs Merged 8/17 Lijian Zhang
Add support for shuffle vector intrinsic via Neon in ARM Merged 8/1 Sirshak Das
Improve cpu { coremask-% } configure option Merged 8/1 Yi He
Fix undefined symbol: fformat_append_cr in vat plugins loading Merged 7/31 Yi He
pp2: increase recycle batch size Merged 7/10 Brian Brooks
pp2: change default queue size Merged 7/26 Brian Brooks
pp2: use configured RX queue size Merged 7/10 Brian Brooks
Fix load_unaligned undefined and other possible build failures Merged 6/26 Sirshak Das
Enable PMU cycle counter for graph node cycles Sirshak Das
Fix clang compilation on aarch64: extraneous parentheses Merged 6/13 Sirshak Das
Fix clang compilation on aarch64: value size does not match register size Merged 5/30 Sirshak Das
Fix clang compilation on aarch64: sizeof operator error Merged 5/30 Sirshak Das
Fix clang compilation on aarch64: replace -pie with -fPIE for dpdk compilation Merged 5/30 Sirshak Das
dpdk: set dmamap iova address value according to eal_iova_mode Merged 5/28 Sachin Saxena
Fixes make test errors with clang compiler on aarch64 Merged 5/27 Sirshak Das
Fix broken compilation for non-numa aware platforms Merged 5/16 Sachin Saxena
build-data: Common makefile for NXP DPAA1/DPAA2 platforms Merged 5/4 Sachin Saxena
arm64: Avoid setting march to corei7 when Cross Compiling for ARM Merged 5/4 Sachin Saxena
use restrict keyword VPP-1126 Khemendra Kumar
Autotools: Autodetection of cache line size VPP-1064 Nitin Saxena
add 'is_all_zero(x)' for NEON - fix build break Merged 2/20 Adrian Oanca
u8x16_compare_byte_mask optimization Merged 2/24 Adrian Oanca
Added u8x16,u32x4,u64x2 variants of _zero_byte_mask(x) for ARM/NEON platform Merged 2/26 VPP-1129 Adrian Oanca
add CLIB_HAVE_VEC128 with NEON intrinsics Merged 02/08 VPP-1127 Gabriel Ganne
Use neutral vector code for ethernet_frame_is_tagged Merged 2/19 Damjan Marion
vhost: Added ARMV8 NEON version of function map_guest_mem() Merged 2/7 VPP-1085 Nitin Saxena
vppinfra: use __atomic_fetch_add instead of __sync_fetch_and_add builtins VPP-1114 Kevin Wang
Arm system counter cleanup Merged 1/30 VPP-1125 Brian Brooks
svm: ... on autodetected VA space size (fixup again) Merged 01/10 Gabriel Ganne
svm: calc base address on AArch64 based on autodetected VA space size (fixup) Merged 01/10 Gabriel Ganne
svm: calc base address on AArch64 based on autodetected VA space size Merged 01/09 Damjan Marion
show cpu microarchitecture Merged 01/06 Gabriel Ganne
Fix Debian Packaging on AARCH64 Merged 01/06 Nitin Saxena
more extended tests fixes Merged 12/16 Gabriel Ganne
Use crc32 wrapper Merged 12/16 VPP-1086 Gabriel Ganne
implement clib_smp_pause() for arm and aarch64 platform Merged 12/15 VPP-1066 Kevin Wang
make "test-all" target pass again (for all platforms) Merged 12/13 Gabriel Ganne
fill "show cpu" Flag list on aarch64 platforms Merged 12/06 VPP-1065 Gabriel Ganne
remove smp dead code Merged 12/06 VPP-1066 Gabriel Ganne
net/virtio: support modern device id Merged 11/28 Gabriel Ganne
use REV on aarch64 for endianness swapping Merged 11/21 VPP-1067 Gabriel Ganne
armv8 crc32 - fix macro name Merged 11/15 Gabriel Ganne
bier - fix node table declaration Merged 11/14 Gabriel Ganne
Map SVM regions at a sane offset on arm64 Merged 11/10 Brian Brooks
bfd tests fix Merged 11/07 Gabriel Ganne
debian packaging fix Merged 11/06 Gabriel Ganne
lb test fix Merged 10/31 Gabriel Ganne
conditional x86intrin.h inclusion Merged 10/25 Gabriel Ganne
fix test_lb_ip4_gre6() cleanup Merged 10/24 Gabriel Ganne
null-terminate some formatted string Merged 10/20 Gabriel Ganne
lb plugin - fix format() type mismatches Merged 10/16 Gabriel Ganne
Use AESNI=y only on x86_64 machines Merged 10/14 Brian Brooks
Improved arm64 chip detection Merged 09/11 Brian Brooks
Native arm64 build: dpdk/Makefile change Merged 08/31 Brian Brooks

Meeting Minutes

11/30/2021

  • Attendees
    • Tianyu Li
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • AVF interface creation issue:
        • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
          • Related to PF i40e driver, vlan stripping configured on VF, PF driver return not allowd.
          • Not related to iavf driver, AVF interface - vpp native driver have this issue
          • dpdk iavf ignore the error and continue initialization, while vpp abort the init process
          • Intel will fix the issue from PF driver - workaround: use old i40e driver (from ubuntu 20.04)
        • Race condition occur on /dev/vfio mounting
            • Run six rounds of tests with patch https://gerrit.fd.io/r/c/csit/+/34045 and result looks good right now
              • Patch has been merged
              • By not mounting whole /dev/vfio, mount only /dev/vfio/xxx used
              • Addressed comments, waiting Peter's review - Peter approved, patch merged and monitoring.
              • Met tx2 server reboot issue when monitoring - RAS CONTROLLER: Fatal unrecoverable error detected ** NBU Error **
              • Reinstall the tx2 server with older kernel version and VPP device testing works fine with Juraj's patch
              • VPP Device configuration align with VPP Performance configuration - no issue yet
          • Enable VPP device testing per patch
            • Monitor for a week and enable vote right then
            • Ping Dave about enabling VPP device testing per patch
    • New Arm servers shipment to the FD.io lab
      • New servers are in the procurement process
      • Plan to replace old thunderx1 build servers with more advanced Arm servers
        • Need to confirm with RAM/disk size for the new build servers
        • 500G disk/256G RAM
        • Each job will consume about 16G memory
      • Intel NIC firmware upgrade on Arm - not supported
  • VPP
    • VPP SVE implementation - Lijian
      • SVE validation on FPGA platform - Conflunence page ready
      • Run L3 traffic testing with NEON/SVE-VLA/SVE-VLS VPP version
        • Performance number: NEON>SVE-VLA>SVE-VLS on FPGA without DMC-620
        • FPGA team promises to provide FPGA image with DMC-620
        • Liangxing will help to benchmark VPP with FPGA enabled with DMC-620
    • VPP IPv4 fragmentation - Tianyu & Jieqiang
      • Add multi-arch support for ip4-frag node but see no perf bump
      • Apply loop unrolling on ip4-frag node
    • VPP IPv6 Benchmarking and Profiling - Jieqiang
      • IPv6 profiling
        • No perf bump for lookup_x2 function in Fd.io gerrit
        • Try Mellaonx nics for IPv6 routing tests
    • CNF PoC proposal preparation- Tianyu
      • Add support for VPP aarch64 docker image build
      • Calico use cases exploration on VPP
      • Improve ansible scripts to deploy VPP&snort on K8S pods automatically
    • VPP IPsec on Arm - Govind
    • VPP Perfmon plugin enablement on Arm - Zach

11/23/2021

  • Attendees
    • Tianyu Li
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • AVF interface creation issue:
        • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
          • Related to PF i40e driver, vlan stripping configured on VF, PF driver return not allowd.
          • Not related to iavf driver, AVF interface - vpp native driver have this issue
          • dpdk iavf ignore the error and continue initialization, while vpp abort the init process
          • Intel will fix the issue from PF driver - workaround: use old i40e driver (from ubuntu 20.04)
        • Race condition occur on /dev/vfio mounting
            • Run six rounds of tests with patch https://gerrit.fd.io/r/c/csit/+/34045 and result looks good right now
              • Patch has been merged
              • By not mounting whole /dev/vfio, mount only /dev/vfio/xxx used
              • Addressed comments, waiting Peter's review - Peter approved, patch merged and monitoring.
              • Met tx2 server reboot issue when monitoring - RAS CONTROLLER: Fatal unrecoverable error detected ** NBU Error **
              • Reinstall the tx2 server with older kernel version and VPP device testing works fine with Juraj's patch
              • VPP Device configuration align with VPP Performance configuration - no issue yet
          • Enable VPP device testing per patch
            • Monitor for a week and enable vote right then
    • New Arm servers shipment to the FD.io lab
      • New servers are in the procurement process
      • Plan to replace old thunderx1 build servers with more advanced Arm servers
        • Need to confirm with RAM/disk size for the new build servers
      • Intel NIC firmware upgrade on Arm - not supported
  • VPP
    • VPP SVE implementation - Lijian
      • SVE validation on FPGA platform - Conflunence page ready
      • Run L3 traffic testing with NEON/SVE-VLA/SVE-VLS VPP version
        • Performance number: NEON>SVE-VLA>SVE-VLS on FPGA without DMC-620
        • FPGA team promises to provide FPGA image with DMC-620
        • Liangxing will help to benchmark VPP with FPGA enbaled with DMC-620
    • VPP IPv4 fragmentation
    • VPP IPv6 Benchmarking and Profiling
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Performance degradation with quad loop unrolling applied on ip6_lookup_inline
        • Patch the current kernel to enable perfmon plugin on VPP
        • Need to check performance for IPv6 subnet routing
        • Try Mellaonx nics for IPv6 routing tests
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
          • Improve ansible scripts to deploy VPP&snort on K8S pods automatically
    • VPP IPsec on Arm - Govind
    • VPP Perfmon plugin enablement on Arm - Zach


11/16/2021

  • Attendees
    • Tianyu Li
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • AVF interface creation issue:
        • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
          • Related to PF i40e driver, vlan stripping configured on VF, PF driver return not allowd.
          • Not related to iavf driver, AVF interface - vpp native driver have this issue
          • dpdk iavf ignore the error and continue initialization, while vpp abort the init process
          • Intel will fix the issue from PF driver - workaround: use old i40e driver (from ubuntu 20.04)
        • Race condition occur on /dev/vfio mounting
            • Run six rounds of tests with patch https://gerrit.fd.io/r/c/csit/+/34045 and result looks good right now
              • Patch has been merged
              • By not mounting whole /dev/vfio, mount only /dev/vfio/xxx used
              • Addressed comments, waiting Peter's review - Peter approved, patch merged and monitoring.
              • Met tx2 server reboot issue when monitoring - RAS CONTROLLER: Fatal unrecoverable error detected ** NBU Error **
              • Reinstall the tx2 server with older kernel version and VPP device testing works fine with Juraj's patch
          • Enable VPP device testing per patch
    • New Arm servers shipment to the FD.io lab
      • New servers are in the procurement process
      • Plan to replace old thunderx1 build servers with more advanced Arm servers
      • Intel NIC firmware upgrade on Arm - not supported
  • VPP
    • VPP SVE implementation - Lijian
      • SVE validation on FPGA platform - Conflunence page ready
      • Run L3 traffic testing with NEON/SVE-VLA/SVE-VLS VPP version
        • Performonce number: NEON>SVE-VLA>SVE-VLS on FPGA without DMC-620
        • FPGA team promises to provide FPGA image with DMC-620
        • Liangxing will help to benchmark VPP with FPGA enbaled with DMC-620
    • VPP IPv4 fragmentation
    • VPP IPv6 Benchmarking and Profiling
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Performance degradation with quad loop unrolling applied on ip6_lookup_inline
        • Patch the current kernel to enable perfmon plugin on VPP
        • Need to check performance for IPv6 subnet routing
        • Try Mellaonx nics for IPv6 routing tests
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
          • Improve ansible scripts to deploy VPP&snort on K8S pods automatically
    • VPP IPsec on Arm - Govind
    • VPP Perfmon plugin enablement on Arm - Zach

11/09/2021

  • Attendees
    • Tianyu Li
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tina Tsou
  • VPP
    • VPP SVE implementation - Lijian
      • SVE validation on FPGA platform - Conflunce page ready
      • Run L3 traffic testing with NEON/SVE-VLA/SVE-VLS VPP version
        • Perfmonce number: NEON>SVE-VLA>SVE-VLS on FPGA without DMC-620
        • FPGA team promises to provide FPGA image with DMC-620
      • VPP IPv4 fragmetation
    • VPP IPv6 Benchmarking and Profiling
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Performance degradation with quad loop unrolling applied on ip6_lookup_inline
        • Patch the current kernel to enable perfmon plugin on VPP
        • Need to check performance for IPv6 subnet routing
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
          • Improve ansible scripts to deploy VPP&snort on K8S pods automatically
    • VPP IPsec on Arm - Govind
    • VPP Perfmon plugin enablement on Arm - Zach

11/02/2021

  • Attendees
    • Tianyu Li
    • Govindarajan Mohandoss
    • Jieqiang Wang
    • Tina Tsou
  • VPP
    • VPP SVE implementation - Lijian
      • SVE validation on FPGA platform - Conflunce page ready
      • Run L3 traffic testing with NEON/SVE-VLA/SVE-VLS VPP version
        • Perfmonce number: NEON>SVE-VLA>SVE-VLS on FPGA without DMC-620
        • FPGA team promises to provide FPGA image with DMC-620
    • VPP IPv6 Benchmarking and Profiling
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Performance degradation with quad loop unrolling applied on ip6_lookup_inline
        • Patch the current kernel to enable perfmon plugin on VPP
        • Need to check performance for IPv6 subnet routing
    • VPP Prefetch
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
          • Improve ansible scripts to deploy VPP&snort on K8S pods automatically
    • VPP IPsec on Arm - Govind
    • VPP Perfmon plugin enablement on Arm - Zach


10/26/2021

  • Attendees
    • Juraj Linkes
    • Tianyu Li
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week. - closed
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • Inbound IPsec: reproduced and need to investigate - Juraj
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • IPsec SPD input/output case ongoing
          • Adding IPsec SPD outbound test cases 64B 1, 100 and 1k SPD entries, 1, 2, 4 cores, on tx2 testbed - clarified
            • Flow cache on and off cases need to be measured.
          • L2 BD 20k test cases execute time too long, removed on taishan.
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
        • 3n-tsh testbed unreachable, investigating right now - Juraj
          • TG firmware is under upgradation
          • Server unreachable due to firmware & driver update - resolved - update all done
        • Release testing for 21.10 starts
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
          • Still see /dev/vfio resource busy error after linux kernel upgradation, but less frequently than before
          • Dig into the log for more details - Juraj
          • Narrow down whether some of those arguments are the reason behind this and that is indeed the case: --volume /dev/vfio:/dev/vfio causes the issue.
        • AVF interface creation issue:
          • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
            • race condition occur
            • try mounting a part of /dev/vfio to see if issue can be resolved
            • Run six rounds of tests with patch https://gerrit.fd.io/r/c/csit/+/34045 and looks good right now
              • Addressed comments, waiting Peter's review..
              • Will enable voting right soon after the patch gets merged
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
        • Lijian has summarized feedbacks from Juraj and raised Jira ticket to DevOps team
        • New servers are in the procurement process
        • Plan to replace old thunderx1 build servers with more advanced Arm servers
            • CPU not fully utilized on Arm, need further investigation
    • Intel NIC firmware upgrade on Arm - not supported
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
            • Run standalone SVE test cases on FPGA
            • Ask for DMC 620 images to run for FPGA
            • Enable DMC 620 more close to real system, but performance will drop
            • Build a system using VPP memif and pktgen
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
        • VPP CLI configuration and hotspot function are recorded in Confluence page
        • Plan to try quad loop unrolling for ip6_lookup_inline function
    • VPP Prefetch
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
          • Try to use ansible to deploy VPP automatically
    • VPP IPsec on Arm - Govind
    • VPP Perfmon plugin enablement on Arm - Zach
      • Review perfmon code - Lijian & Govind
      • Implemented statistics from PMUv3 - done
        • Patch upstream has dependency on kernel patch, waiting for this before upstream to VPP
        • Building Intel QAT driver on arm to test IPsec crypt - Zach

10/19/2021

  • Attendees
    • Juraj Linkes
    • Tianyu Li
    • Govindarajan Mohandoss
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
          • Still see /dev/vfio resource busy error after linux kernel upgradation, but less frequently than before
          • Dig into the log for more details - Juraj
          • Narrow down whether some of those arguments are the reason behind this and that is indeed the case: --volume /dev/vfio:/dev/vfio causes the issue.
        • AVF interface creation issue:
          • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
            • race condition occur
            • try mounting a part of /dev/vfio to see if issue can be resolved
            • Run six rounds of tests with patch https://gerrit.fd.io/r/c/csit/+/34045 and looks good right now
              • Will enable voting right soon after the patch gets merged
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
        • Lijian has summarized feedbacks from Juraj and raised Jira ticket to DevOps team
        • New servers are in the procurement process
        • Plan to replace old thunderx1 build servers with more advanced Arm servers
            • CPU not fully utilized on Arm, need further investigation
    • Intel NIC firmware upgrade on Arm - not supported
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
            • Run standalone SVE test cases on FPGA
            • Ask for DMC 620 images to run for FPGA
            • Enable DMC 620 more close to real system, but performance will drop
            • Build a system using VPP memif and pktgen
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
        • VPP CLI configuration and hotspot function are recorded in Confluence page
        • Plan to try quad loop unrolling for ip6_lookup_inline function
    • VPP Prefetch
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
          • Try to use ansible to deploy VPP automatically
    • VPP IPsec on Arm - Govind
    • VPP Perfmon plugin enablement on Arm - Zach
      • Review perfmon code - Lijian & Govind
      • Implemented statistics from PMUv3 - done
        • Patch upstream has dependency on kernel patch, waiting for this before upstream to VPP
        • Building Intel QAT driver on arm to test IPsec crypt - Zach

10/12/2021

  • Attendees
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Govindarajan Mohandoss
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
          • Still see /dev/vfio resource busy error after linux kernel upgradation, but less frequently than before
          • Dig into the log for more details - Juraj
          • Narrow down whether some of those arguments are the reason behind this and that is indeed the case: --volume /dev/vfio:/dev/vfio causes the issue.
        • AVF interface creation issue:
          • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
            • race condition occur
            • try mounting a part of /dev/vfio to see if issue can be resolved
            • Talked with Peter, Juraj is working on prototype of mounting part of /dev/vfio
            • x86 vpp device job is fine, duo to firmware & driver is old
            • arm vpp device servers have drivers updated, vlan striping not allowed, vlan configuration cannot removed from lab view.
            • only performance testbeds have NIC drivers updated
            • maintainer doesn't want to a option from vpp config
            • may need to check x86 have the same issue with the same version driver before reaching intel folks
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
        • Lijian has summarized feedbacks from Juraj and raised Jira ticket to DevOps team
        • New servers are in the procurement process
        • Plan to replace old thunderx1 build servers with more advanced Arm servers
            • CPU not fully utilized on Arm, need further investigation
    • Intel NIC firmware upgrade on Arm - not supported
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
            • Run standalone SVE test cases on FPGA
            • Ask for DMC 620 images to run for FPGA
            • Enable DMC 620 more close to real system, but performance will drop
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
        • VPP CLI configuration and hotspot function are recorded in Confluence page
    • VPP Prefetch
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • Discussion on the default action on the IPsec inbound interface which does not match
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach
      • Review perfmon code - Lijian & Govind
      • Implemented statistics from PMUv3 - done
        • Patch upstream has dependency on kernel patch, waiting for this before upstream to VPP
        • Building Intel QAT driver on arm to test IPsec crypt - Zach

09/28/2021

  • Attendees
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
          • Still see /dev/vfio resource busy error after linux kernel upgradation, but less frequently than before
          • Dig into the log for more details - Juraj
          • Narrow down whether some of those arguments are the reason behind this and that is indeed the case: --volume /dev/vfio:/dev/vfio causes the issue.
        • AVF interface creation issue:
          • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
            • race condition occur
            • try mounting a part of /dev/vfio to see if issue can be resolved
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
        • Lijian has summarized feedbacks from Juraj and raised Jira ticket to DevOps team
        • New servers are in the procurement process
        • Plan to replace old thunderx1 build servers with more advanced Arm servers
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
            • Run standalone SVE test cases on FPGA
            • Ask for DMC 620 images to run for FPGA
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
        • VPP CLI configuration and hotspot function are recorded in Confluence page
    • VPP Prefetch
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • Discussion on the default action on the IPsec inbound interface which does not match
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach
      • Review perfmon code - Lijian & Govind
      • Implemented statistics from PMUv3 - done
        • Patch upstream has dependency on kernel patch, waiting for this before upstream to VPP
        • Building Intel QAT driver on arm to test IPsec crypt - Zach

09/14/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
            • Try VPP IPsec test cases with a fix on outbound interface - Govind & Juraj
              • Outbound IPsec: 10 entries still slower with latest change, related to traffic pattern
              • Outbound IPsec finished.
              • Waiting for new version of patchset to verify test cases
            • Will try the fix on inbound IPsec tests when the Jenkins is back to normal - Juraj
              • Inbound IPsec: reproduced and need to investigate - Juraj
                • Learn more about RFC and need time to understand more - on hold - waiting Neale's response
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • Release testing done.
          • A few more jobs run for release 21.06 and will be finished soon
          • Comparison between 21.06 and 21.01.1 is ongoing.
        • IPsec SPD input/output case ongoing
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
          • Still see /dev/vfio resource busy error after linux kernel upgradation, but less frequently than before
          • Dig into the log for more details - juraj
          • Narrow down whether some of those arguments are the reason behind this and that is indeed the case: --volume /dev/vfio:/dev/vfio causes the issue.
        • AVF interface creation issue:
          • Can't create AVF interface on VFs with configured VLAN - happens with latest i40e driver on tx2
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
        • Lijian has summarized feedbacks from Juraj and raised Jira ticket to DevOps team
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
            • Run standalone SVE test cases on FPGA
            • Ask for DMC 620 images to run for FPGA
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
        • VPP CLI configuration and hotspot function are recorded in Confluence page
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
      • Try IPv4 multicast testing to verify the scenario when refcnt > 1
        • GDB shows that mbufs are copied instead of reference from src port to all dst ports
        • Direct/Indirect mbuf for VPP multicast testing
        • Try IPv4 multicasting & L2 flood testing which works fine
        • ip4-replicate node in IPv4 multicasting/l2-flood node in L2 flood testing
          • show mbuf is copied so that ref_cnt will always be one
            • dpdk 21.08 have the patches, need to verify on vpp
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • Discussion on the default action on the IPsec inbound interface which does not match
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach
      • Review perfmon code - Lijian & Govind
      • Implemented statistics from PMUv3 - done
        • Patch upstream has dependency on kernel patch, waiting for this before upstream to VPP
        • Building Intel QAT driver on arm to test IPsec crypt - Zach

09/07/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
            • Try VPP IPsec test cases with a fix on outbound interface - Govind & Juraj
              • Outbound IPsec: 10 entries still slower with latest change, related to traffic pattern
              • Outbound IPsec finished.
              • Waiting for new version of patchset to verify test cases
            • Will try the fix on inbound IPsec tests when the Jenkins is back to normal - Juraj
              • Inbound IPsec: reproduced and need to investigate - Juraj
                • Learn more about RFC and need time to understand more - on hold - waiting Neale's response
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • Release testing done.
          • A few more jobs run for release 21.06 and will be finished soon
          • Comparison between 21.06 and 21.01.1 is ongoing.
        • IPsec SPD input/output case ongoing
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
          • Still see /dev/vfio resource busy error after linux kernel upgradation, but less frequently than before
          • Dig into the log for more details - juraj
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
        • Lijian has summarized feedbacks from Juraj and raised Jira ticket to DevOps team
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
            • Run standalone SVE test cases on FPGA
            • Ask for DMC 620 images to run for FPGA
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
        • VPP CLI configuration and hotspot function are recorded in Confluence page
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
      • Try IPv4 multicast testing to verify the scenario when refcnt > 1
        • GDB shows that mbufs are copied instead of reference from src port to all dst ports
        • Direct/Indirect mbuf for VPP multicast testing
        • Try IPv4 multicasting & L2 flood testing which works fine
        • ip4-replicate node in IPv4 multicasting/l2-flood node in L2 flood testing
          • show mbuf is copied so that ref_cnt will always be one
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • Discussion on the default action on the IPsec inbound interface which does not match
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach
      • Review perfmon code - Lijian & Govind
      • Implemented statistics from PMUv3 - done
        • Patch upstream has dependency on kernel patch, waiting for this before upstream to VPP
        • Building Intel QAT driver on arm to test IPsec crypt - Zach

08/31/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
            • Try VPP IPsec test cases with a fix on outbound interface - Govind & Juraj
              • Outbound IPsec: 10 entries still slower with latest change, related to traffic pattern
              • Outbound IPsec finished.
              • Waiting for new version of patchset to verify test cases
            • Will try the fix on inbound IPsec tests when the Jenkins is back to normal - Juraj
              • Inbound IPsec: reproduced and need to investigate - Juraj
                • Learn more about RFC and need time to understand more - on hold - waiting Neale's response
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • Release testing done.
          • A few more jobs run for release 21.06 and will be finished soon
          • Comparison between 21.06 and 21.01.1 is ongoing.
        • IPsec SPD input/output case ongoing
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
        • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
          • config option set to N, /dev/vfio device or resource busy error
          • config option set to Y & iommu_passthrough = 1, IP packets Rx timeout
            • The longer the server runs, more test cases fails
          • Next to do
            • Need to figure out what arm-smmu-v3.4.auto: event 0x10 means
              • Juraj modified script to reproduce the issue - Lijian will try it locally
              • Also seen in Intel QAT card from Zach
            • Will try to reproduce this issue on local thx2 with 20.04 distro - Lijian
            • Lijian have slight different firmware version, driver version
            • Tried Mellanox card (rdma driver) multiple times - not see the same issue (on XL710 NIC) happens - Juraj
            • Lijian can use Juraj's script to reproduce the issue on local tx2 server
              • Reducing the numa buffer allocation size resolves this issue
              • Observed from the error log of numa buffer allocation
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
            • Run standalone SVE test cases on FPGA
            • Ask for DMC 620 images to run for FPGA
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
        • VPP CLI configuration and hotspot function are recorded in Confluence page
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
      • Try IPv4 multicast testing to verify the scenario when refcnt > 1
        • GDB shows that mbufs are copied instead of reference from src port to all dst ports
        • Direct/Indirect mbuf for VPP multicast testing
        • Try IPv4 multicasting & L2 flood testing which works fine
        • ip4-replicate node in IPv4 multicasting/l2-flood node in L2 flood testing
          • show mbuf is copied so that ref_cnt will always be one
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • Discussion on the default action on the IPsec inbound interface which does not match
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach
      • Review perfmon code - Lijian & Govind
      • Implemented statistics from PMUv3 - done

08/24/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
            • Try VPP IPsec test cases with a fix on outbound interface - Govind & Juraj
              • Outbound IPsec: 10 entries still slower with latest change, related to traffic pattern
              • Outbound IPsec finished.
              • Waiting for new version of patchset to verify test cases
            • Will try the fix on inbound IPsec tests when the Jenkins is back to normal - Juraj
              • Inbound IPsec: reproduced and need to investigate - Juraj
                • Learn more about RFC and need time to understand more - on hold - waiting Neale's response
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • Release testing done.
          • A few more jobs run for release 21.06 and will be finished soon
          • Comparison between 21.06 and 21.01.1 is ongoing.
        • IPsec SPD input/output case ongoing
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
        • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
          • config option set to N, /dev/vfio device or resource busy error
          • config option set to Y & iommu_passthrough = 1, IP packets Rx timeout
            • The longer the server runs, more test cases fails
          • Next to do
            • Need to figure out what arm-smmu-v3.4.auto: event 0x10 means
              • Juraj modified script to reproduce the issue - Lijian will try it locally
              • Also seen in Intel QAT card from Zach
            • Will try to reproduce this issue on local thx2 with 20.04 distro - Lijian
            • Lijian have slight different firmware version, driver version
            • Tried Mellanox card (rdma driver) multiple times - not see the same issue (on XL710 NIC) happens - Juraj
            • Lijian can use Juraj's script to reproduce the issue on local tx2 server
              • Reducing the numa buffer allocation size resolves this issue
              • Observed from the error log of numa buffer allocation
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
        • Need vexxhost guys confirm with ethernet / power cable type info.
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
          • Able to access FPGA platform, investigating adding vpp to buildroot filesystem - Lijian
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
      • Try IPv4 multicast testing to verify the scenario when refcnt > 1
        • GDB shows that mbufs are copied instead of reference from src port to all dst ports
        • Will try L2 flood test case & understand VPP/multicast code
        • Direct/Indirect mbuf for VPP multicast testing
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
        • Damjan has merged 2 patches, waiting for the last patch, for generic 128B cacheline size.
        • For 64B cacheline size native build on Arm, may need to change code.
        • Current VPP does not support 64B cacheline size compilation for Arm images.
        • Issues about prefetch on current VPP code base
          • Issue 1 support 128B/64B cache-line size in Arm image
          • Issue 2 prefetch 'overflow' for native build
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • Discussion on the default action on the IPsec inbound interface which does not match
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach

08/17/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
            • Try VPP IPsec test cases with a fix on outbound interface - Govind & Juraj
              • Outbound IPsec: 10 entries still slower with latest change, related to traffic pattern
              • Outbound IPsec finished.
              • Waiting for new version of patchset to verify test cases
            • Will try the fix on inbound IPsec tests when the Jenkins is back to normal - Juraj
              • Inbound IPsec: reproduced and need to investigate - Juraj
                • Learn more about RFC and need time to understand more
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • Release testing ongoing
          • A few more jobs run for release 21.06 and will be finished soon
          • Comparison between 21.06 and 21.01.1 is ongoing.
        • IPsec SPD input/output case ongoing
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
        • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
          • config option set to N, /dev/vfio device or resource busy error
          • config option set to Y & iommu_passthrough = 1, IP packets Rx timeout
            • The longer the server runs, more test cases fails
          • Next to do
            • Need to figure out what arm-smmu-v3.4.auto: event 0x10 means
              • Juraj modified script to reproduce the issue - Lijian will try it locally
              • Also seen in Intel QAT card from Zach
            • Will try to reproduce this issue on local thx2 with 20.04 distro - Lijian
            • Lijian have slight different firmware version, driver version
            • Tried Mellanox card (rdma driver) multiple times - not see the same issue (on XL710 NIC) happens - Juraj
            • Lijian can use Juraj's script to reproduce the issue on local tx2 server
              • Reducing the numa buffer allocation size resolves this issue
              • Observed from the error log of numa buffer allocation
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
    • VPN access request to FD.io Arm servers
      • Lijian/Jieqiang has got VPN access now
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • IPv6 profiling
        • Hotspot function - ip6_lookup_node/ip6_rewrite_node
        • Will try perfmon & understand two node functions
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
      • Try IPv4 multicast testing to verify the scenario when refcnt > 1
        • GDB shows that mbufs are copied instead of reference from src port to all dst ports
        • Will try L2 flood test case & understand VPP/multicast code
        • Direct/Indirect mbuf for VPP multicast testing
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
        • Damjan has merged 2 patches, waiting for the last patch, for generic 128B cacheline size.
        • For 64B cacheline size native build on Arm, may need to change code.
        • Current VPP does not support 64B cacheline size compilation for Arm images.
        • Issues about prefetch on current VPP code base
          • Issue 1 support 128B/64B cache-line size in Arm image
          • Issue 2 prefetch 'overflow' for native build
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • Discussion on the default action on the IPsec inbound interface which does not match
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach

08/10/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Test cases with 1, 10, 100, 1000 SPD entries - still under review
            • Try VPP IPsec test cases with a fix on outbound interface - Govind & Juraj
              • Outbound IPsec: 10 entries still slower with latest change, related to traffic pattern
              • Outbound IPsec finished.
              • Waiting for new version of patcheset to verify test cases
            • Will try the fix on inbound IPsec tests when the Jenkins is back to normal - Juraj
              • Inbound IPsec: reproduced and need to investigate
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • Release testing ongoing
          • Comparison between 21.06 and 21.01.1 is ongoing.
        • IPsec SPD input/output case ongoing
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863

`

    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
        • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
          • config option set to N, /dev/vfio device or resource busy error
          • config option set to Y & iommu_passthrough = 1, IP packets Rx timeout
            • The longer the server runs, more test cases fails
          • Next to do
            • Need to figure out what arm-smmu-v3.4.auto: event 0x10 means
              • Juraj modified script to reproduce the issue - Lijian will try it locally
              • Also seen in Intel QAT card from Zach
            • Will try to reproduce this issue on local thx2 with 20.04 distro - Lijian
            • Lijian have slight different firmware version, driver version
            • Tried Mellanox card (rdma driver) multiple times - not see the same issue (on XL710 NIC) happens - Juraj
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
    • VPN access request to FD.io Arm servers
      • Lijian has got VPN access now
      • Juraj singed Jieqiang's key
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128, CLI issue only, CSIT's python API works fine.
      • Internal patch to resolve this issue under review - upstreamed
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
        • Damjan has merged 2 patches, waiting for the last patch, for generic 128B cacheline size.
        • For 64B cacheline size native build on Arm, may need to change code.
        • Current VPP does not support 64B cacheline size compilation for Arm images.
    • VPP Classifier - Lijian
      • 3 patches: prefetch, key-value compare simd improvement, cache to look up
      • 4 loop unrolling decreasing performance
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach

08/03/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2106-3n-tsh/
        • 2106 testing partial finished. 21.01.1 ongoing, should be done sometime next week.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Test cases with 1, 10, 100, 1000 SPD entries - still under review
            • Try VPP IPsec test cases with a fix on outbound interface - Govind & Juraj
              • Outbound IPsec: 10 entries still slower with latest change, related to traffic pattern
              • Waiting for new version of patcheset to verify test cases
            • Will try the fix on inbound IPsec tests when the Jenkins is back to normal - Juraj
              • Inbound IPsec: reproduced and need to investigate
          • Flow cache with 1, 10 SPD entries slower, still investigating. Manual test local vs CSIT have different result on 1-10 SPD policies.
        • Release testing ongoing
          • Comparison between 21.06 and 21.01.1 is ongoing.
        • IPsec SPD input/output case ongoing
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863

`

    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
      • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
        • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
          • Resulting in the same failure as before, only happen on AArch64 platform
        • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
          • config option set to N, /dev/vfio device or resource busy error
          • config option set to Y & iommu_passthrough = 1, IP packets Rx timeout
            • The longer the server runs, more test cases fails
          • Next to do
            • Need to figure out what arm-smmu-v3.4.auto: event 0x10 means
              • Also seen in Intel QAT card from Zach
            • Will try to reproduce this issue on local thx2 with 20.04 distro - Lijian
            • Will try Mellanox card to see if same issue happens - Juraj
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
    • VPN access request to FD.io Arm servers
      • Lijian has got VPN access now
      • Juraj singed Jieqiang's key
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
    • VPP IPv6 Benchmarking and Profiling
      • VPP CLI 'ip route add ipv6_addr/mask' outputs wrong IPv6 routes with mask 123-128
      • Internal patch to resolve this issue under review
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
        • Damjan has merged 2 patches, waiting for the last patch, for generic 128B cacheline size.
        • For 64B cacheline size native build on Arm, may need to change code.
        • Current VPP does not support 64B cacheline size compilation for Arm images.
    • VPP Classifier - Lijian
      • 3 patches: prefetch, key-value compare simd improvement, cache to look up
      • 4 loop unrolling decreasing performance
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
          • Calico use cases exploration on VPP
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach

07/27/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
          • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
            • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
            • Resulting in the same failure as before, only happen on AArch64 platform
            • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
            • Random issue, more frequently happening on Arm
              • Not see in CI recently or manually.
        • scapy unexpected timeout issue: packet drop or slow issue?
            • vfio-pci driver may be the root cause - bind/unbind
      • Connection issue between Jenkins and the build executor in FD.io lab
    • Shipment of new advanced server to the FD.io lab
      • One link between TG and DUT, multiple link between DUT for testing LACP.
      • Two advanced servers are in plan to ship
    • VPN access request to FD.io Arm servers
      • Lijian has got VPN access now
      • Juraj singed Jieqiang's key
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
          • Rebased SVE patch per Nitin's request, waiting for Nitin's feedback on running these patches
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform - Conflunce page ready
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
    • VPP mbuf-fast-free tx offload
      • Vector path shows performance improvement, still need to investigate scalar path
      • Performance improvement for IPv4 routing test cases using vector path
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
        • Damjan has merged 2 patches, waiting for the last patch, for generic 128B cacheline size.
        • For 64B cacheline size native build on Arm, may need to change code.
    • VPP Classifier - Lijian
      • 3 patches: prefetch, key-value compare simd improvement, cache to look up
      • 4 loop unrolling descreasing performance
    • VPP memif - Tianyu
      • CNF PoC proposal preparation
          • Add support for VPP aarch64 docker image build
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch - merged
    • VPP Perfmon plugin enablement on Arm - Zach

07/20/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated.
            • This is fixed in DPDK 21.05 version by making iavf PMD as default.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
          • New issue: different error related moving VF from host to container(not involving dpdk/VPP) - just started investigating
            • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
            • Resulting in the same failure as before, only happen on AArch64 platform
            • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
            • Random issue, more frequently happening on Arm
            • vfio-pci driver may be the root cause - bind/unbind
      • Connection issue between Jenkins and the build executor in FD.io lab
    • Shipment of new advanced server to the FD.io lab
      • Two advanced servers are in plan to ship
    • VPN access request to FD.io Arm servers
      • Lijian has got VPN access now
      • Juraj singed Jieqiang's key
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform
          • Run unit tests from DPDK and VPP bihash on FPGA
          • Try Lijian's SVE patch to see any cycle count improvement
    • VPP mbuf-fast-free tx offload
      • Performance improvement for IPv4 routing test cases using vector path
    • VPP Prefetch
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP Classifier - Lijian
    • VPP memif - Tianyu
      • Investigating VPP memif - Tianyu
    • VPP IPsec on Arm - Govind
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
      • SPD prototype change on ipsec_output/encryption node - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch
    • VPP Perfmon plugin enablement on Arm - Zach

07/13/2021

  • Attendees
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Test cases with 1, 10, 100, 1000 SPD entries - still under review
            • Expected to be merged soon
          • Flow cache with 1, 10 SPD entries slower, still investigating. Mannual test local vs CSIT have different result on 1-10 SPD policies.
            • Hugepage size, numa-node, core isolation etc. may need to check.
          • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - Fixed
        • Release testing ongoing
        • IPsec SPD input/output case ongoing
        • Juraj may share the steps how CSIT handle new configuration changes
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Some container test cases failed on all platform. - 4 cores container case fail on x86 and arm - Fixed and passing.
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
        • 21.06 vs 21.01 see performance drop on https://docs.fd.io/csit/master/report/_static/vpp/performance-changes-3n-tsh-1t1c-pdr.txt
          • May need to check VM and IPsec cases
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated. After using iavf PMD, the issue is not seen. This is fixed in DPDK 21.05 version by making iavf PMD as default. This fix will be ported to all the older VPP LTS release branches. Currently, the fix is planned only for VPP 20.09 release.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
          • New issue: different error related moving VF from host to containter(not involving dpdk/VPP) - just started investigating
            • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
            • Resulting in the same failure as before, only happen on AArch64 platform
            • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
            • Random issue, more frequently happening on Arm
            • vfio-pci driver may be the root cause - bind/unbind
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
    • Shipment of new advanced server to the FD.io lab
      • New servers are in shortage.
    • VPN access request to FD.io Arm servers
      • Will remind Machiek to sign Lijian's GPG public key.
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)


07/06/2021

  • Attendees
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Test cases with 1, 10, 100, 1000 SPD entries - still under review
            • Expected to be merged soon
          • Flow cache with 1, 10 SPD entries slower, still investigating. Mannual test local vs CSIT have different result on 1-10 SPD policies.
            • Hugepage size, numa-node, core isolation etc. may need to check.
          • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - Fixed
        • Release testing ongoing
        • IPsec SPD input/output case ongoing
        • Juraj may share the steps how CSIT handle new configuration changes
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Some container test cases failed on all platform. - 4 cores container case fail on x86 and arm - Fixed and passing.
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
        • 21.06 vs 21.01 see performance drop on https://docs.fd.io/csit/master/report/_static/vpp/performance-changes-3n-tsh-1t1c-pdr.txt
          • May need to check VM and IPsec cases
    • VPP Path
      • Voting and working fine.
      • CentOS-8 jobs have been removed.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated. After using iavf PMD, the issue is not seen. This is fixed in DPDK 21.05 version by making iavf PMD as default. This fix will be ported to all the older VPP LTS release branches. Currently, the fix is planned only for VPP 20.09 release.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
          • New issue: different error related moving VF from host to containter(not involving dpdk/VPP) - just started investigating
            • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
            • Resulting in the same failure as before, only happen on AArch64 platform
            • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
            • Random issue, more frequently happening on Arm
            • vfio-pci driver may be the root cause - bind/unbind
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
    • Shipment of new advanced server to the FD.io lab
      • New servers are in shortage.
    • VPN access request to FD.io Arm servers
      • Will remind Machiek to sign Lijian's GPG public key.
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform
    • VPP Prefetch
      • Benchmark VPP using prefetch read always vs prefetch write always on N1SDP - Jieqiang
        • Repeat the same test on Ampere server - PMU cache-miss less for write always
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP Classifier - Lijian
      • 3 patches: prefetch, key-value compare simd improvement, cache to look up
      • Investigating VPP classify function, use case, benchmarking - Lijian
        • Start with simple use case
        • VPP Classify basic inbound L3 src ip / prot case
        • Benchmark VPP classifier on Arm/X86 platform
          • Done some NEON changes, see some microbranchmark improvement
      • investigate CSIT case
        • No classify test case in CSIT. - Jieqiang
        • may be there is a CSIT case named iacldstbase
    • VPP memif - Tianyu
      • Investigating VPP memif - Tianyu
        • Benchmarking DPDK memif vs VPP memif
          • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
          • Performance improvement using loop unrolling for memif nodes - running CSIT perftest
          • Patches have been upstreamed and waiting for review
          • Prepare the memif readout - Tianyu
          • Add support for VPP aarch64 docker image build
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
      • Memif C11 atomics has been updated by maintainer, not using atomic_relaxed - Tianyu
    • VPP IPsec on Arm - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
        • VPP uses linear search on SPD lookups
      • SPD prototype change on ipsec_output/encryption node, introducing flow cache with hash, has performance improvements, discussing with community - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch
    • VPP Perfmon plugin enablement on Arm - Zach

06/29/2021

  • Attendees
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Test cases with 1, 10, 100, 1000 SPD entries - still under review
            • Expected to be merged soon
          • Flow cache with 1, 10 SPD entries slower, still investigating. Mannual test local vs CSIT have different result on 1-10 SPD policies.
            • Hugepage size, numa-node, core isolation etc. may need to check.
          • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - Fixed
        • Release testing ongoing
        • IPsec SPD input/output case ongoing
        • Juraj may share the steps how CSIT handle new configuration changes
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Some container test cases failed on all platform. - 4 cores container case fail on x86 and arm - Fixed and passing.
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
        • 21.06 vs 21.01 see performance drop on https://docs.fd.io/csit/master/report/_static/vpp/performance-changes-3n-tsh-1t1c-pdr.txt
          • May need to check VM and IPsec cases
    • VPP Path
      • Voting and working fine.
      • Community plans to drop the support for CentOS-8.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated. After using iavf PMD, the issue is not seen. This is fixed in DPDK 21.05 version by making iavf PMD as default. This fix will be ported to all the older VPP LTS release branches. Currently, the fix is planned only for VPP 20.09 release.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
          • New issue: different error related moving VF from host to containter(not involving dpdk/VPP) - just started investigating
            • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
            • Resulting in the same failure as before, only happen on AArch64 platform
            • /sys/module/vfio/parameters/enable_unsafe_noiommu_mode affect the behavior in some way
            • Debugging
            • vfio-pci driver may be the root cause - bind/unbind
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
    • Shipment of new advanced server to the FD.io lab
      • New servers are in shortage.
    • VPN access request to FD.io Arm servers
      • mbuf-fast-free dpdk enablement with VPP (DEV_TX_OFFLOAD_MBUF_FAST_FREE)
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform
    • VPP Prefetch
      • Benchmark VPP using prefetch read always vs prefetch write always on N1SDP - Jieqiang
        • Repeat the same test on Ampere server - PMU cache-miss less for write always
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP Classifier - Lijian
      • 3 patches: prefetch, key-value compare simd improvement, cache to look up
      • Investigating VPP classify function, use case, benchmarking - Lijian
        • Start with simple use case
        • VPP Classify basic inbound L3 src ip / prot case
        • Benchmark VPP classifier on Arm/X86 platform
          • Done some NEON changes, see some microbranchmark improvement
      • investigate CSIT case
        • No classify test case in CSIT. - Jieqiang
        • may be there is a CSIT case named iacldstbase
    • VPP memif - Tianyu
      • Investigating VPP memif - Tianyu
        • Benchmarking DPDK memif vs VPP memif
          • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
          • Performance improvement using loop unrolling for memif nodes - running CSIT perftest
          • Patches have been upstreamed and waiting for review
          • Prepare the memif readout - Tianyu
          • Add support for VPP aarch64 docker image build
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
      • Memif C11 atomics has been updated by maintainer, not using atomic_relaxed - Tianyu
    • VPP IPsec on Arm - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
        • VPP uses linear search on SPD lookups
      • SPD prototype change on ipsec_output/encryption node, introducing flow cache with hash, has performance improvements, discussing with community - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Upstreamed https://gerrit.fd.io/r/c/vpp/+/32903
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch
    • VPP Perfmon plugin enablement on Arm - Zach

06/22/2021

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Test cases with 1, 10, 100, 1000 SPD entries
            • Expected to be merged soon
          • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - Fixed
        • Release testing ongoing
        • IPsec SPD input/output case ongoing
        • Juraj may share the steps how CSIT handle new configuration changes
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Some container test cases failed on all platform. - 4 cores container case fail on x86 and arm
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
    • VPP Path
      • Voting and working fine.
      • Community plans to drop the support for CentOS-8.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated. After using iavf PMD, the issue is not seen. This is fixed in DPDK 21.05 version by making iavf PMD as default. This fix will be ported to all the older VPP LTS release branches. Currently, the fix is planned only for VPP 20.09 release.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
          • New issue: different error related moving VF from host to containter(not involving dpdk/VPP) - just started investigating
            • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
            • Resulting in the same failure as before, only happen on AArch64 platform
            • vfio-pci driver may be the root cause
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
    • Shipment of new advanced server to the FD.io lab
      • New servers are in shortage.
    • VPN access request to FD.io Arm servers
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform
    • VPP Prefetch
      • Benchmark VPP using prefetch read always vs prefetch write always on N1SDP - Jieqiang
        • Repeat the same test on Ampere server
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP Classifier - Lijian
      • Investigating VPP classify function, use case, benchmarking - Lijian
        • Start with simple use case
        • VPP Classify basic inbound L3 src ip / prot case
        • Benchmark VPP classifier on Arm/X86 platform
          • Done some NEON changes, see some microbranchmark improvement
      • investigate CSIT case
        • No classify test case in CSIT. - Jieqiang
        • may be there is a CSIT case named iacldstbase
    • VPP memif - Tianyu
      • Investigating VPP memif - Tianyu
        • Benchmarking DPDK memif vs VPP memif
          • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
          • Performance improvement using loop unrolling for memif nodes - running CSIT perftest
          • Patches have been upstreamed and waiting for review
          • Prepare the memif readout - Tianyu
          • Add support for VPP aarch64 docker image build
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
      • Memif C11 atomics has been updated by maintainer, not using atomic_relaxed - Tianyu
    • VPP IPsec on Arm - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
        • VPP uses linear search on SPD lookups
      • SPD prototype change on ipsec_output/encryption node, introducing flow cache with hash, has performance improvements, discussing with community - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Internal review for IPsec input node flow cache implementation - Zach & Govind
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch
    • VPP Perfmon plugin enablement on Arm - Zach

06/15/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
          • VPP exits with the IPsec startup config, try startup config from Zach's email - Juraj
          • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - Fixed
        • Release testing ongoing
        • IPsec SPD input/output case ongoing
        • Juraj may share the steps how CSIT handle new configuration changes
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Some container test cases failed on all platform. - 4 cores container case fail on x86 and arm
        • Steps to enable test case in CSIT https://gerrit.fd.io/r/c/csit/+/31863
    • VPP Path
      • Voting and working fine.
      • Community plans to drop the support for CentOS-8.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated. After using iavf PMD, the issue is not seen. This is fixed in DPDK 21.05 version by making iavf PMD as default. This fix will be ported to all the older VPP LTS release branches. Currently, the fix is planned only for VPP 20.09 release.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results - not reproduced so far.
          • New issue: different error related moving VF from host to containter(not involving dpdk/VPP) - just started investigating
            • /usr/bin/vpp[3789]: pci: 0000:91:02.0: open_vfio_iommu_group: open '/dev/vfio/141': Device or resource busy
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly. - DaveW
    • Shippment of new adavanced server to the FD.io lab
      • New servers are in shortage.
  • VPP
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
        • Juan met and fixing some issue running SVE in qemu VM
        • SVE validation on FPGA platform
    • VPP Prefetch
      • Benchmark VPP using prefetch read always vs prefetch write always on N1SDP - Jieqiang
        • Repeat the same test on Ampere server
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP Classifier - Lijian
      • Investigating VPP classify function, use case, benchmarking - Lijian
        • Start with simple use case
        • VPP Classify basic inbound L3 src ip / prot case
        • Benchmark VPP classifier on Arm/X86 platform
          • Done some NEON changes, see some microbranchmark improvement
      • investigate CSIT case
        • No classify test case in CSIT. - Jieqiang - may be there is a CSIT case named iacldstbase
    • VPP memif - Tianyu
      • Investigating VPP memif - Tianyu
        • Benchmarking DPDK memif vs VPP memif
          • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
          • Performance improvement using loop unrolling for memif nodes - running CSIT perftest
          • Patches have been upstreamed and waiting for review
          • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
      • Memif C11 atomics has been updated by maintainer, not using atomic_relaxed - Tianyu
    • VPP IPsec on Arm - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input/output nodes - Govind & Zach - waiting for maintainer review
        • VPP uses linear search on SPD lookups
      • SPD prototype change on ipsec_output/encryption node, introducing flow cache with hash, has performance improvements, discussing with community - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Waiting for review comments on outbound side before upstream to VPP
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch
    • VPP Perfmon plugin enablement on Arm - Zach


06/08/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Add new IPSec NULL encryption & decryption test cases - Juraj
          • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
          • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
          • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
          • VPP exits with the IPsec startup config, try startup config from Zach's email - Juraj
          • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - Work in progress.
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Some container test cases failed on all platform.
    • VPP Path
      • Voting and working fine.
      • Community plans to drop the support for CentOS-8.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated. After using iavf PMD, the issue is not seen. This is fixed in DPDK 21.05 version by making iavf PMD as default. This fix will be ported to all the older VPP LTS release branches. Currently, the fix is planned only for VPP 20.09 release.
          • Use iavf PMD instead of i40evf on all VPP branches, waiting for the test results.
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
    • Shippment of new adavanced server to the FD.io lab
      • New servers are in shortage.
  • VPP
    • VPP default compiler on Arm platform
      • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
        • Plan to benchmark gcc-10 vs clang-12 on arm L2/L3, 1/10k - Jieqiang
          • No obvious performance improvement, keep the original default compiler
    • VPP SVE implementation - Lijian
      • Vector length specific patch is ready
      • SVE patch ready and upstreamed, under review - Lijian
        • SVE patch sent to Nitin, Nitin will review the patch when back to work.
        • SVE validation on FPGA platform
    • VPP Prefetch
      • Benchmark VPP using prefetch read always vs prefetch write always - Jieqiang
      • Refactor prefetch implementation in VPP per CPU's actual cache line size - Tianyu & Jieqiang
    • VPP Classifier - Lijian
      • Investigating VPP classify function, use case, benchmarking - Lijian
        • Start with simple use case
        • VPP Classify basic inbound L3 src ip / prot case
        • Benchmark VPP classifier on Arm/X86 platform
      • investigate CSIT case
        • No classify test case in CSIT. - Jieqiang
    • VPP memif - Tianyu
      • Investigating VPP memif - Tianyu
        • Benchmarking DPDK memif vs VPP memif
          • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
          • Performance improvement using loop unrolling for memif nodes - running CSIT perftest
          • Patches have been upstreamed and waiting for review
          • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
      • Memif C11 atomics has been updated by maintainer, not using atomic_relaxed - Tianyu
    • VPP IPsec on Arm - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input/output nodes - Govind & Zach
        • VPP uses linear search on SPD lookups
      • SPD prototype change on ipsec_output/encryption node, introducing flow cache with hash, has performance improvements, discussing with community - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Waiting for review comments on outbound side before upstream to VPP
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch
    • VPP Perfmon plugin enablement on Arm - Zach


06/01/2021

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Zachary Leaf
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
          • Start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • IPSec policy test cases are not running by default.
            • Some of the IPSec test cases(Policy tests) has been added to daily testing.
            • Enabled the policy tests in mrr-daily testing and it's now running on both 2n-tx2 and 3n-tsh (and also available for per-patch on-demand testing) - Juraj
            • Add new IPSec NULL encryption & decryption test cases - Juraj
              • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
              • IPSec SPD test cases will be ready next week, how to make SPD policy change - IP address range changes - Juraj
            • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
            • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - Work in progress.
            • Send the correct robot framework tags for IPSec policy test cases to Govind - Juraj
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Juraj will investigate cabling issue on Taishan performance test-bed - resolved.
          • Some container case are seems failure on all platform.
    • VPP Path
      • Voting and working fine.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
          • Internal ticket has been raised
            • DPDK uses i40evf PMD and it is old and scheduled to be deprecated. After using iavf PMD, the issue is not seen. This is fixed in DPDK 21.05 version by making iavf PMD as default. This fix will be ported to all the older VPP LTS release branches. Currently, the fix is planned only for VPP 20.09 release.
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
      • Plan to benchmark gcc-10 vs clang-12 on arm L2/L3, 1/10k - Jieqiang
    • Vector length specific patch is ready
    • Investigating VPP classify function, use case, benchmarking - Lijian
      • Start with simple use case
      • 128 and 256 fixed size vector wrappers are ready, needs verification
    • Jieqiang helped to verify most fixed size vector wrappers - unit test code
    • SVE Remaining works - variable type convention - need some workaround for 256bit convention
    • VLA patch, coding and verification done. string memcpy/memset, bihash key compare functions, rdma/bonding node
    • SVE patch ready and upstreamed, under review - Lijian
      • SVE patch sent to Nitin, Nitin will review the patch when back to work.
      • Review memif patch
        • Memif C11 atomics has been updated by maintainer, not using atomic_relaxed - Tianyu
      • VPP Classify basic inbound L3 src ip / prot case, investigate CSIT case - No classify test case in CSIT. - Jieqiang
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
        • Review the confluence page and prepare the memif readout - Lijian & Tianyu
        • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
        • Performance improvement using loop unrolling for memif nodes - running CSIT perftest
        • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
        • Run memif unit test
    • IPsec on Arm platform - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
    • Work on IPsec input/output nodes - VPP uses linear search on SPD lookups - Govind & Zach
      • SPD prototype change on ipsec_output/encryption node, introducing flow cache with hash, has performance improvements, discussing with community - Govind
        • https://gerrit.fd.io/r/c/vpp/+/31694
        • IPSec unit tests done - 'make test' cases implemented & included in patch - Zach
          • Testing of flow cache functionality, including hash collisions and stale entry overwrites
      • IPSec input node/decryption flow cache implemented in a separate patch - Zach
        • Waiting for review comments on outbound side before upstream to VPP
        • Discovered issue with SPD policy counter/statistics on input side, to be fixed in additional standalone patch
    • Perfmon plugin enablement on Arm - Zach

05/25/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Zachary Leaf
    • Tianyu Li
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
          • Start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • IPSec policy test cases are not running by default.
            • Some of the IPSec test cases(Policy tests) has been added to daily testing.
            • Enabled the policy tests in mrr-daily testing and it's now running on both 2n-tx2 and 3n-tsh (and also available for per-patch on-demand testing) - Juraj
            • Add new IPSec NULL encryption & decryption test cases - Juraj
              • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
              • IPSec SPD test cases will be ready next week, how to make SPD policy change - IP address range changes - Juraj
            • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
            • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308 - patch fully tested, waiting for review - Juraj
            • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready - will look into it
            • Send the correct robot framework tags for IPSec policy test cases to Govind - Juraj
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Juraj will investigate cabling issue on Taishan performance test-bed - resolved.
          • Some container case are seems failure on all platform.
    • VPP Path
      • Voting and working fine.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
            • Intel folks debugged, tried updating firmware/drivers, VF driver updated, old driver: i40evf, new driver: iavf. After modified dpdk code by using iavf, issue can be fixed. Need to find proper solution of dpdk.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
          • Internal ticket has been raised
            • Try the new version of DPDK but it does not help
            • Contact Intel devs for the possible advice
            • Workaround may impact too much to all test cases
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
      • Plan to benchmark gcc-10 vs clang-12 on arm L2/L3, 1/10k - Jieqiang
    • Vector length specific patch is ready
    • Investigating VPP classify function, use case, benchmarking - Lijian
      • Start with simple use case
      • 128 and 256 fixed size vector wrappers are ready, needs verification
    • Jieqiang helped to verify most fixed size vector wrappers - unit test code
    • SVE Remaining works - variable type convention - need some workaround for 256bit convention
    • VLA patch, coding and verification done. string memcpy/memset, bihash key compare functions, rdma/bonding node
    • SVE patch ready and upstreamed, under review - Lijian
      • SVE patch sent to Nitin, Nitin will review the patch when back to work.
      • Review memif patch
        • Memif C11 atomics has been updated by maintainer, not using atomic_relaxed - Tianyu
      • VPP Classify basic inbound L3 src ip / prot case, investigate CSIT case - No classify test case in CSIT. - Jieqiang
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
        • Review the confluence page and prepare the memif readout - Lijian & Tianyu
        • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
        • Performance improvement using loop unrolling for memif nodes - running CSIT perftest
        • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
        • Run memif unit test
    • IPsec on Arm platform. - Govind
      • IPsec input node optimization work in progress - Zach & Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
          • https://gerrit.fd.io/r/c/vpp/+/31694
          • IPSec unit test - make test new cases implementation
          • Make test cases for IPSec policy mode - Done, included in Govind's patch, waiting for maintainer review - Zach
            • Add/Remove/Add+Remove+Readd/Hash collisions/Multiple interfaces & rules
          • Review the patch and grasp the basics about IPSec - Lijian
        • Perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind
    • Working on IPsec flow cache, discussed with Neal, maintainer agree with the change
    • perfmon CMN-600 investigating - Zach
      • VPP perfmon CMN-600 patch abandon, system level, not vpp node level, linux perf can give the same result - Zach
      • Plan to upstream perfmon plugin - resolving review comments - Zach
      • IPSec flow cache outbound done, working on inbound side in seperate patch - Zach
      • IPSec decryption / input node - Zach

05/18/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Zachary Leaf
    • Tianyu Li
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
          • Start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • IPSec policy test cases are not running by default.
            • Some of the IPSec test cases(Policy tests) has been added to daily testing.
            • Enabled the policy tests in mrr-daily testing and it's now running on both 2n-tx2 and 3n-tsh (and also available for per-patch on-demand testing) - Juraj
            • Add new IPSec NULL encryption & decryption test cases - Juraj
              • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
              • IPSec SPD test cases will be ready next week, how to make SPD policy change - IP address range changes - Juraj
            • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
            • New IPsec test cases - https://gerrit.fd.io/r/c/csit/+/32308
            • Enable flow cache option in startup.conf for VPP CSIT IPsec test cases when the patch is ready
            • Send the correct robot framework tags for IPSec policy test cases to Govind - Juraj
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Juraj will investigate adding IPSec test cases on Taishan performance test-bed.
    • VPP Path
      • Voting and working fine.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
          • Try to reproduce with another set of firmware and etc but issues still exist
          • https://doc.dpdk.org/guides/nics/i40e.html
          • Internal ticket has been raised
            • Try the new version of DPDK but it does not help
            • Contact Intel devs for the possible advice
            • Workaround may impact too much to all test cases
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
      • Lab move for the fd.io lab
        • Lab moving started stage 2, moved part of the servers to make sure ci service not down.
        • Lab move is done, some issues with taishan testbed
        • Arm servers documented in https://gerrit.fd.io/r/c/csit/+/30662
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
      • Plan to benchmark gcc-10 vs clang-12
    • Vector length specific patch is ready
      • 128 and 256 fixed size vector wrappers are ready, needs verification
    • Jieqiang helped to verify most fixed size vector wrappers - unit test code
    • SVE Remaining works - variable type convention - need some workaround for 256bit convention
    • VLA patch, coding and verification done. string memcpy/memset, bihash key compare functions, rdma/bonding node
    • SVE patch ready and upstreamed, under review - Lijian
      • SVE patch sent to Nitin, Nitin will review the patch when back to work.
      • Review memif patch
        • Functional bug related to C11 atomics has been resolved by VPP maintainer.
      • VPP Classify basic inbound L3 src ip / prot case, investigate CSIT case. - Jieqiang
      • Make test cases for IPSec policy mode - Zach
        • Add/Remove/Add+Remove+Readd/Hash collisions/Multiple interfaces & rules - Add more test cases
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
        • Review the confluence page and prepare the memif readout - Lijian & Tianyu
        • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
        • Performance improvement using loop unrolling for memif nodes
        • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
        • Run memif unit test
    • IPsec on Arm platform. - Govind
      • IPsec input node optimization work in progress - Zach & Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
        • Perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind
    • Working on IPsec flow cache, discussed with Neal, maintainer agree with the change
    • perfmon CMN-600 investigating - Zach
      • VPP perfmon CMN-600 patch abandon, system level, not vpp node level, linux perf can give the same result - Zach
      • Plan to upstream perfmon plugin - resolving review comments - Zach
      • IPSec flow cache outbound done, working on inbound side in seperate patch - Zach
      • IPSec decryption / input node - Zach

05/11/2021

  • Attendees
    • Lijian Zhang
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Zachary Leaf
    • Tianyu Li
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
          • Start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • IPSec policy test cases are not running by default.
            • Some of the IPSec test cases(Policy tests) has been added to daily testing.
            • Enabled the policy tests in mrr-daily testing and it's now running on both 2n-tx2 and 3n-tsh (and also available for per-patch on-demand testing) - Juraj
            • Add new IPSec NULL encryption & decryption test cases - Juraj
              • New IPSec SPD test cases will not have NULL encrypt/decrypt config.
              • IPSec SPD test cases will be ready next week, how to make SPD policy change - IP address range changes - Juraj
            • CSIT Maintainers approved to add new IPSec Policy mode test cases with multiple SPD rules.
            • Send the correct robot framework tags for IPSec policy test cases to Govind - Juraj
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Juraj will investigate adding IPSec test cases on Taishan performance test-bed.
    • VPP Path
      • Voting and working fine.
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • Moved the NIC from arm server to x86 server can easily update firmware, given script to reproduce, intel busy with other stuff, debugging in progress.
          • https://doc.dpdk.org/guides/nics/i40e.html
          • Internal ticket has been raised
            • Try the new version of DPDK but it does not help
            • Contact Intel devs for the possible advice
            • Workaround may impact too much to all test cases
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
      • Lab move for the fd.io lab
        • Lab moving started stage 2, moved part of the servers to make sure ci service not down.
        • Arm servers documented in https://gerrit.fd.io/r/c/csit/+/30662
        • Almost all except performance testbed, which will be moved this week, everything is smooth so far.
        • ubuntu 1804 -> 2004
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Vector length specific patch is ready
      • 128 and 256 fixed size vector wrappers are ready, needs verification
    • Jieqiang helped to verify most fixed size vector wrappers - unit test code
    • SVE Remaining works - variable type convention - need some workaround for 256bit convention
    • VLA patch, coding and verification done. string memcpy/memset, bihash key compare functions, rdma/bonding node
    • SVE patch ready and upstreamed, under review - Lijian
      • SVE patch sent to Nitin, Nitin will review the patch when back to work.
      • Review memif patch
      • VPP Classify basic inbound L3 src ip / prot case, investigate CSIT case.
      • Make test cases for IPSec policy mode - Jieqiang
    • Review memif test cases/memif cases
    • Finished coding of SVE string library, bihash key compare functions
    • SVE unit testing based on test_vec, fix test_vec issues
    • Investigate VPP test cases in container
      • Focus more on data-plane performance benchmarking and optimization - Tianyu
        • Record the benchmarking results of VPP CNF 3 test cases in excel template
          • Will do readout presentation with extended people - Tianyu
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
        • Review the confluence page and prepare the memif readout - Lijian & Tianyu
        • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
        • Performance improvement using loop unrolling for memif nodes
        • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
        • Run memif unit test
    • IPsec on Arm platform. - Govind
      • IPsec input node optimization work in progress - Zach & Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
        • Perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind
    • Working on IPsec flow cache, discussed with Neal, maintainer agree with the change
    • Using startup parameter to enable the IPsec flow cache feature
    • Adding Python test case to test IPSec node behavior - Jieqiang
    • perfmon CMN-600 investigating - Zach
      • VPP perfmon CMN-600 patch abandon, system level, not vpp node level, linux perf can give the same result - Zach
      • Plan to upstream perfmon plugin - resolving review comments - Zach
      • IPSec flow cache outbound done, working on inbound side in seperate patch - Zach
      • IPSec decryption / input node - Zach

04/27/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
          • Start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • IPSec policy test cases are not running by default.
            • Some of the IPSec test cases(Policy tests) has been added to daily testing.
            • Enabled the policy tests in mrr-daily testing and it's now running on both 2n-tx2 and 3n-tsh (and also available for per-patch on-demand testing) - Juraj
            • Add new IPSec NULL encryption & decryption test cases - Juraj
            • Send the correct robot framework tags for IPSec policy test cases to Govind - Juraj
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Juraj will investigate adding IPSec test cases on Taishan performance test-bed.
    • VPP Path
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
          • https://doc.dpdk.org/guides/nics/i40e.html
          • Internal ticket has been raised
            • Try the new version of DPDK but it does not help
            • Contact Intel devs for the possible advice
            • Workaround may impact too much to all test cases
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Vector length specific patch is ready
      • 128 and 256 fixed size vector wrappers are ready, needs verification
    • Jieqiang helped to verify most fixed size vector wrappers - unit test code
    • SVE Remaining works - variable type convention - need some workaround for 256bit convention
    • VLA patch, coding and verification done. string memcpy/memset, bihash key compare functions, rdma/bonding node
    • SVE patch ready and upstreamed, under review - Lijian
      • Make test cases for IPSec policy mode - Jieqiang
    • Review memif test cases/memif cases
    • Finished coding of SVE string library, bihash key compare functions
    • SVE unit testing based on test_vec, fix test_vec issues
    • Investigate VPP test cases in container
      • Focus more on data-plane performance benchmarking and optimization - Tianyu
        • Record the benchmarking results of VPP CNF 3 test cases in excel template
          • Will do readout presentation with extended people - Tianyu
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
        • Review the confluence page and prepare the memif readout - Lijian & Tianyu
        • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
        • Performance improvement using loop unrolling for memif nodes
        • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
        • Run memif unit test
    • IPsec on Arm platform. - Govind
      • IPsec input node optimization work in progress - Zach & Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
        • Perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind
    • Working on IPsec flow cache, discussed with Neal, maintainer agree with the change
    • Using startup parameter to enable the IPsec flow cache feature
    • Adding Python test case to test IPSec node behavior - Jieqiang
    • perfmon CMN-600 investigating - Zach
      • Plan to upstream perfmon plugin - resolving review comments - Zach
      • IPSec decryption / input node - Zach

04/13/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
          • Start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • IPSec policy test cases are not running by default.
            • Some of the IPSec test cases(Policy tests) has been added to daily testing.
            • Enabled the policy tests in mrr-daily testing and it's now running on both 2n-tx2 and 3n-tsh (and also available for per-patch on-demand testing) - Juraj
            • Add new IPSec NULL encryption & decryption test cases - Juraj
            • Send the correct robot framework tags for IPSec policy test cases to Govind - Juraj
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Some issues occurred during the upgrade.
          • Patch to resolve the building error of DPDK on 3n-tsh testbed.
          • Root cause is the change of build system of DPDK on 3n-tsh related to SOC id detection.
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Juraj will investigate adding IPSec test cases on Taishan performance test-bed.
    • VPP Path
    • VPP Device
      • https://jenkins.fd.io/view/vpp/job/vpp-csit-verify-device-master-ubuntu2004-aarch64-1n-tx2/
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
        • Will try to reproduce the issue with x86 servers.
        • This issue is common to all platforms(Arm & Intel)
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Vector length specific patch is ready
      • 128 and 256 fixed size vector wrappers are ready, needs verification
    • Jieqiang helped to verify most fixed size vector wrappers - unit test code
    • SVE Remaining works - variable type convention - need some workaround for 256bit convention
    • VLA patch, coding and verification done. string memcpy/memset, bihash key compare functions, rdma/bonding node
      • Make test cases for IPSec policy mode - Jieqiang
    • Review memif test cases/memif cases
    • Finished coding of SVE string library, bihash key compare functions
    • SVE unit testing based on test_vec, fix test_vec issues
    • Test template update - Jieqiang
    • Investigate VPP test cases in container
      • Focus more on data-plane performance benchmarking and optimization - Tianyu
        • Record the benchmarking results of VPP CNF 3 test cases in excel template
          • Will do readout presentation with extended people - Tianyu
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
        • Review the confluence page and prepare the memif readout - Lijian & Tianyu
        • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
        • Prepare the memif readout - Tianyu
      • Try to apply C11 weak memory model on VPP memif - Tianyu
        • Use 'show runtime'/perfmon to see cycle improvement
        • Run memif unit test
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
        • Perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind
    • Working on IPsec flow cache, discussed with Neal, maintainer agree with the change
    • Using startup parameter to enable the IPsec flow cache feature
    • Adding Python test case to test IPSec node behavior - Jieqiang
    • perfmon CMN-600 investigating - Zach
      • Plan to upstream perfmon plugin - resolving review comments - Zach
      • IPSec decryption - Zach


03/30/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
          • Start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • IPSec policy test cases are not running by default.
            • 2 node IPsec SPD policy test case patch is ready, starting with 1 and 1k tunnels. (40, 400 tunnels in seperate patch)
            • https://gerrit.fd.io/r/c/csit/+/31605
            • Fix the wrong CLI commands but configuration still has problems.
            • Send the correct robot framework tags for IPSec policy test cases to Govind - Juraj
        • 2n-tx2 & 3n-tsh has been upgraded to ubuntu 20.04, everything is working fine now.
          • Some issues occurred during the upgrade.
          • Patch to resolve the building error of DPDK on arm testbed.(taishan dpdk cases still have issues, investigating)
          • Juraj is investigating running those test cases with 2N-TX2 topology.
          • Juraj will investigate adding IPSec test cases on Taishan performance test-bed.
    • VPP Path
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • The issue could be reproduced on Arm servers with the NIC with latest firmware version. (still reproduced, no update yet)
        • Will try to reproduce the issue with x86 servers.
      • "Make Test Testcase Error or Failure" --> There was an intermittent VPP anomaly introduced last week with the change from shmem to socket transport for PAPI which causes the MAP unittests to fail [0]. The root cause is being addressed and should be fixed shortly.
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Vector length specific patch is ready
      • 128 and 256 fixed size vector wrappers are ready, needs verification
    • Review memif test cases/memif cases
    • Finished coding of SVE string library, bihash key compare functions
    • SVE unit testing based on test_vec, fix test_vec issues
    • Test template update
    • SVE unit test in qemu-vm, met compiling issue, investigating
    • Investigate VPP test cases in container
      • Focus more on data-plane performance benchmarking and optimization - Tianyu
        • Record the benchmarking results of VPP CNF 3 test cases in excel template
          • Will do readout presentation with extended people - Tianyu
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
        • Review the confluence page and prepare the memif readout - Lijian & Tianyu
        • Race condition occurred hen connecting DPDK memif PMD interface(slave) with VPP memif interface(master)
        • Prepare the memif readout - Tianyu
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
        • Perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind
    • Working on IPsec flow cache, discussed with Neal, maintainer agree with the change
    • Using startup parameter to enable the IPsec flow cache feature
    • Discuss with jieqiang adding python test case to test ipsec node behavior
    • perfmon CMN-600 investigating - Zach

03/16/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Vector length specific patch is ready
      • 128 and 256 fixed size vector wrappers are ready, needs verification
      • Verify SVE vector length specific wrappers - Jieqiang
      • https://gerrit.fd.io/r/c/vpp/+/30326 - second proposal - not preferred
        • Extend vector length agnostic opportunities
    • Investigate VPP test cases in container
      • Focus more on data-plane performance benchmarking and optimization - Tianyu
        • Record the benchmarking results of VPP CNF 3 test cases in excel template
          • Will do readout presentation with extented people - Tianyu
    • Investigating VPP memif - Tianyu
      • Benchmarking DPDK memif vs VPP memif
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
        • Perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

03/09/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • CentOS-7 will be enabled with master branch for support lts release
        • CentOS-7 Jenkins on Arm will not be supported.
        • CentOS-8 will be supported by the end of this year by Redhat.
      • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge'
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
        • Dependency on maintainers to fix this issue
        • Intel will ship a new NIC with latest firmware
          • Shipment takes a long time empirically
            • NIC has been shipped to vexxhost, wait for NIC arrival.
          • Try to reproduce the issue on this NIC on Arm platform
          • Updating firmware on the current NIC is risky
        • Voting rights will be enabled once this issue is fixed
            • Maintainer raised the ticket to get intel people involved
            • Will not update the firmware because the release testing is ongoing
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maintainers
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
        • Will show Arm roadmap in the next TSC meeting
        • Proposals have been sent to VPP maintainer on verifying SVE/SVE2
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
        • avf-input node with neon optimization is merged.
        • ethernet-input patch needs to split into two parts required by VPP maintainer
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
      • Try to capture some software benchmarking results
      • https://gerrit.fd.io/r/c/vpp/+/29942 - first proposal - preferred
        • 128 and 256 fixed size vector wrappers are ready, needs verification
        • Verify SVE vector length specific wrappers - Jieqiang
      • https://gerrit.fd.io/r/c/vpp/+/30326 - second proposal - not preferred
        • Extend vector length agnostic opportunities
      • Investigate the scalable SIMD instructions on RISC-V - Lijian
      • Investigate how to run traffic tests for VPP in docker - Lijian
        • Plan to talk with VPP maintainers on this topic
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
        • Analysis of benchmarking results for Ampere Altra
          • A lot of context switches occur on Ampere Altra compared to N1SDP
          • perf tools used to capture the perf events
          • Talk with Ampere or N1 team on how to enable CMN-600 counters for ampere altra
          • Remove interrupts on altra but no performance improvement seen
          • instruction cache misses are higher on altra than N1
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
    • Investigate VPP test cases in container
      • Focus more on data-plane performance benchmarking and optimization - Tianyu
        • Record the benchmarking results of VPP CNF 3 test cases in excel template
    • VPP compiling error on CentOS 7 - Jieqiang
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • SPD prototype change, introducing flow cache with hash, have performance improvement, discussing with community.
        • perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

02/23/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
      • CentOS-7 will be enabled with master branch for support lts release
        • CentOS-7 Jenkins on Arm will be supported.
      • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge'
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
        • Dependency on maintainers to fix this issue
        • Intel will ship a new NIC with latest firmware
          • Shipment takes a long time empirically
          • Try to reproduce the issue on this NIC on Arm platform
          • Updating firmware on the current NIC is risky
        • Voting rights will be enabled once this issue is fixed
            • Maintainer raised the ticket to get intel people involved
            • Will not update the firmware because the release testing is ongoing
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maintainers
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
      • Verify SVE/SVE2 features inside QEMU docker - Lijian
        • Latest VPP binary crash on the QEMU docker
          • System call fails inside QEMU docker when running VPP
        • Verify SVE/SVE2 features inside ARM QEMU VM
        • VPP maintainers want real hardware to verify SVE code
          • This solution will be abandoned.
        • 'make test' execution is slow
        • Sync with DPDK team/VPP community to decide the solution
        • Proposals have been sent to VPP maintainer on verifying SVE/SVE2
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
        • avf-input node with neon optimization is merged.
        • ethernet-input patch needs to split into two parts required by VPP maintainer
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
      • Try to capture some software benchmarking results
      • https://gerrit.fd.io/r/c/vpp/+/29942 - first proposal - preferred
        • 128 and 256 fixed size vector wrappers are ready, needs verification
        • Verify SVE vector length specific wrappers - Jieqiang
      • https://gerrit.fd.io/r/c/vpp/+/30326 - second proposal - not preferred
        • Extend vector length agnostic opportunities
      • Investigate the scalable SIMD instructions on RISC-V - Lijian
      • Investigate how to run traffic tests for VPP in docker - Lijian
        • Plan to talk with VPP maintainers on this topic
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • No positive for 4x loop unrolling on Ampere, so will keep 2x unrolling for Neoverse N1
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
        • Analysis of benchmarking results for Ampere Altra
          • A lot of context switches occur on Ampere Altra compared to N1SDP
          • perf tools used to capture the perf events
          • Talk with Ampere or N1 team on how to enable CMN-600 counters for ampere altra
          • Remove interrupts on altra but no performance improvement seen
          • instruction cache misses are higher on altra than N1
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
    • Investigate VPP test cases in container
      • Investigate VPP memif test cases in VPP CSIT - Jieqiang
      • Investigate VPP use cases proposals in containers - Tianyu
        • 3 use cases are investigated.
        • Will explore the memif logic and share the progress.
        • Will share the link on details about how to run VPP in container.
      • Investigate VPP agent usage - Tianyu
        • Focus more on data-plane performance benchmarking and optimization - Tianyu
    • VPP compiling error on CentOS 7 - Jieqiang
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • perfmon plugin enablement on Arm - Zach
          • patch upstream has dependency on kernel patch
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

02/09/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
        • Dependency on maintainers to fix this issue
        • Intel will ship a new NIC with latest firmware
          • Shippment takes a long time empirically
          • Try to reproduce the issue on this NIC on Arm platform
          • Updating firmware on the current NIC is risky
        • Voting rights will be enabled once this issue is fixed
            • Maintainer raised the ticket to get intel people involved
            • Will not update the firmware because the release testing is ongoing
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maintainers
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
      • Verify SVE/SVE2 features inside QEMU docker - Lijian
        • Latest VPP binary crash on the QEMU docker
          • System call fails inside QEMU docker when running VPP
        • Verify SVE/SVE2 features inside ARM QEMU VM
        • 'make test' execution is slow
        • Sync with DPDK team/VPP community to decide the solution
        • Proposals have been sent to VPP maintainer on verifying SVE/SVE2
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
        • avf-input node with neon optimization is merged.
        • ethernet-input patch needs to split into two parts required by VPP maintainer
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
      • Try to capture some software benchmarking results
      • https://gerrit.fd.io/r/c/vpp/+/29942 - first proposal - preferred
      • https://gerrit.fd.io/r/c/vpp/+/30326 - second proposal - not preferred
      • Investigate the scalable SIMD instructions on RISC-V - Lijian
      • Investigate how to run traffic tests for VPP in docker - Lijian
        • Plan to talk with VPP maintainers on this topic
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • No positive for 4x loop unrolling on Ampere, so will keep 2x unrolling for Neoverse N1
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
        • Analysis of benchmarking results for Ampere Altra
          • A lot of context switches occur on Ampere Altra compared to N1SDP
          • perf tools used to capture the perf events
          • Talk with Ampere or N1 team on how to enable CMN-600 counters for ampere altra
          • Remove interrupts on altra but no performance improvement seen
          • instruction cache misses are higher on altra than N1
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
    • Investigate VPP test cases in container
      • Investigate VPP memif test cases in VPP CSIT - Jieqiang
      • Investigate VPP use cases proposals in containers - Tianyu
        • 3 use cases are investigated.
        • Will explore the memif logic and share the progress.
        • Will share the link on details about how to run VPP in container.
      • Investigate VPP agent usage - Tianyu
    • Benchmark Altra vs Cascade 8268
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

02/02/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
    • VPP Path
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
        • Dependency on maintainers to fix this issue
        • Voting rights will be enabled once this issue is fixed
          • Implementation is ready, and will do test it with actual patches.
          • Apply file locking mechanism to allow that only one VPP instances are running.
            • https://gerrit.fd.io/r/c/csit/+/30425
            • Patches are under review
            • Maintainer raised the ticket to get intel people involved
            • Will not update the firmware because the release testing is ongoing
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maintainers
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
      • Verify SVE/SVE2 features inside QEMU docker - Lijian
        • Latest VPP binary crash on the QEMU docker
          • System call fails inside QEMU docker when running VPP
        • Verify SVE/SVE2 features inside ARM QEMU VM
        • 'make test' execution is slow
        • Sync with DPDK team/VPP community to decide the solution
        • Proposals have been sent to VPP maintainer on verifying SVE/SVE2
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
        • avf-input node with neon optimization is merged.
        • ethernet-input patch needs to split into two parts required by VPP maintainer
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
      • Try to capture some software benchmarking results
      • https://gerrit.fd.io/r/c/vpp/+/29942 - first proposal - preferred
      • https://gerrit.fd.io/r/c/vpp/+/30326 - second proposal - not preferred
      • Investigate the scalable SIMD instructions on RISC-V - Lijian
      • Investigate how to run traffic tests for VPP in docker - Lijian
        • Plan to talk with VPP maintainers on this topic
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • No positive for 4x loop unrolling on Ampere, so will keep 2x unrolling for Neoverse N1
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
        • Analysis of benchmarking results for Ampere Altra
          • A lot of context switches occur on Ampere Altra compared to N1SDP
          • perf tools used to capture the perf events
          • Talk with Ampere or N1 team on how to enable CMN-600 counters for ampere altra
          • Remove interrupts on altra but no performance improvement seen
          • instruction cache misses are higher on altra than N1
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
    • Investigate VPP test cases in container
      • Investigate VPP memif test cases in VPP CSIT - Jieqiang
      • Investigate VPP use cases proposals in containers - Tianyu
        • 3 use cases are investigated.
        • Will explore the memif logic and share the progress.
        • Will share the link on details about how to run VPP in container.
      • Investigate VPP agent usage - Tianyu
    • Benchmark Altra vs Cascade 8268
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

01/19/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tianyu Li
    • Jieqiang Wang
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2101-3n-tsh/
        • CSIT official release 20.09 is available
          • https://docs.fd.io/csit/rls2009/report/
          • Jieqiang will compare the performance data with release 20.09
            • Will investigate test case eth-l2xcbase-eth-2vhostvr1024-1vm - Jieqiang
            • DPDK testpmd running inside VM, l2 cross connect running inside VPP.
            • Check the number for CSIT 2101 release
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Hardware configurations/wiring are done; Physical connection to the TG is done.
        • almost done, two steps need to be done
          • start with basic L2/L3/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • Take the execution time into consideration if we want run release testing on 2n-thx2.
          • It takes 9 hours to finish the one round testing.
          • Tests are running fine
            • L2/L3 tests are running fine, IPSec tests are not supported on 2-node topo, ACL/Classifiers needs investigation.
            • Suitable time to run release testing on 2n-tx2 testbed.
            • Will investigate IPSec test cases on 2n-tx2 - Juraj
            • Add memif test case to 2n-tx2 once the release testing is done.
    • VPP Path
      • CentOS-7 will be enabled with master branch for support lts release
        • CentOS-7 Jenkins on Arm will be supported.
      • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge'
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
          • Implementation is ready, and will do test it with actual patches.
          • Apply file locking mechanism to allow that only one VPP instances are running.
            • https://gerrit.fd.io/r/c/csit/+/30425
            • Patches are under review
            • Machiek raised the ticket to get intel people involved
            • Will not update the firmaware because the release testing is ongoing
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maintainers
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
      • Verify SVE/SVE2 features inside QEMU docker
        • Latest VPP binary crash on the QEMU docker - Lijian
      • Lab move for the fd.io lab
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
        • avf-input node with neon optimization is merged.
        • ethernet-input patch needs to split into two parts required by VPP maintainer
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
      • Try to capture some software benchmarking results
      • https://gerrit.fd.io/r/c/vpp/+/29942 - first proposal - preferred
      • https://gerrit.fd.io/r/c/vpp/+/30326 - second proposal - not preferred
      • Investigate the scalable SIMD instructions on RISC-V - Lijian
      • Investigate how to run traffic tests for VPP in docker - Lijian
        • Plan to talk with VPP maintainers on this topic
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • No positive for 4x loop unrolling on Ampere, so will keep 2x unrolling for Neoverse N1
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
        • Analysis of benchmarking results for Ampere Altra
          • A lot of context switches occur on Ampere Altra compared to N1SDP
          • perf tools used to capture the perf events
          • Talk with Ampere or N1 team on how to enable CMN-600 counters for ampere altra
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
    • Investigate VPP test cases in container
      • Investigate VPP memif test cases in VPP CSIT - Jieqiang
      • Investigate VPP use cases proposals in containers - Tianyu
        • 3 use cases are investigated.
        • Will explore the memif logic and share the progress.
        • Will share the link on details about how to run VPP in container.
    • Benchmark Altra vs Cascade 8268
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind


01/05/2021

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Jieqiang Wang
    • Tianyu Li
    • Tina Tsou
  • CSIT
    • VPP Performance Test
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-mrr-weekly-master-3n-tsh/
      • https://jenkins.fd.io/view/csit/job/csit-dpdk-perf-report-iterative-2009-3n-tsh/
        • CSIT official release 20.09 is available
          • https://docs.fd.io/csit/rls2009/report/
          • Jieqiang will compare the performance data with release 20.09
            • Will investigate test case eth-l2xcbase-eth-2vhostvr1024-1vm - Jieqiang
            • DPDK testpmd running inside VM, l2 cross connect running inside VPP.
      • Leverage current spare TX2 server as 2-node topology performance test-bed.
        • Hardware configurations/wiring are done; Physical connection to the TG is done.
        • almost done, two steps need to be done
          • start with basic L2/L3/IPSec/ACL/classifiers tests and use DPDK PMD on 2n-thx2 firstly(daily testing)
          • Take the execution time into consideration if we want run release testing on 2n-thx2.
          • Tests are running fine
            • L2/L3 tests are running fine, IPSec tests are not supported on 2-node topo, ACL/Classifiers needs investigation.
            • Suitable time to run release testing on 2n-tx2 testbed.
    • VPP Path
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
          • Implementation is ready, and will do test it with actual patches.
          • Apply file locking mechanism to allow that only one VPP instances are running.
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maintainers
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
      • Verify SVE/SVE2 features inside QEMU docker
        • Latest VPP binary crash on the QEMU docker - Lijian
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
        • avf-input node with neon optimization is merged.
        • ethernet-input patch needs to split into two parts required by VPP maintainer
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
      • Try to capture some software benchmarking results
      • https://gerrit.fd.io/r/c/vpp/+/29942 - first proposal - preferred
      • https://gerrit.fd.io/r/c/vpp/+/30326 - second proposal - not preferred
      • Investigate the scalable SIMD instructions on RISC-V - Lijian
      • Investigate how to run traffic tests for VPP in docker - Lijian
        • Plan to talk with VPP maintainers on this topic
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • No positive for 4x loop unrolling on Ampere, so will keep 2x unrolling for Neoverse N1
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
    • Investigate VPP test cases in container
      • Investigate VPP test cases in VPP CSIT - Jieqiang
      • Investigate VPP use cases proposals in containers - Tianyu
    • Benchmark Altra vs Cascade 8268
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

12/22/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
    • Will cancel the meeting on Dec 29th;
  • CSIT
    • VPP Performance Test
    • VPP Path
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maintainers
      • LF will provide QSFP+ fiber switch for FD.io lab.
        • Basically done. LF just procured the existing fiber switch currently rent by Arm in FD.io lab.
        • Send the progress to relavent people in Arm - Lijian
        • Confirm with Tina to ensure Arm is not charged - Lijian
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
      • Verify SVE/SVE2 features on VPP CSIT
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
        • avf-input node with neon optimization is merged.
        • ethernet-input patch needs to split into two parts required by VPP maintainer
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • No positive for 4x loop unrolling on Ampere, so will keep 2x unrolling for Neoverse N1
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
      • Try the VFs with DPDK plugin. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
    • Benchmark Altra vs Cascade 8268
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind


12/15/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
    • Will cancel the meeting on Dec 29th;
  • CSIT
    • VPP Performance Test
    • VPP Path
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • VPP community is responding this issue actively. - Juraj
        • Working on a workaround to make sure not starting multiple VPP instances at the same time - Juraj
          • Implementation is ready, and will do test it with actual patches.
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maitainers
      • LF will provide QSFP+ fiber switch for FD.io lab.
        • Basically done. LF just procured the existing fiber switch currently rent by Arm in FD.io lab.
        • Send the progress to relavent people in Arm - Lijian
      • Arm is required to present Arm achievement and plan to TSC.
        • Govind will prepare the slides
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • SOC ID will be available on /proc entry starting from kernel version 5.9 to differentiate vendor CPUs.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • Optimize ethernet-input and avf-input node with NEON intrinsics
        • Benchmarking result shows some improvement from vectorization with ethernet-input and avf-input node
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals are upstreamed, will discuss the proposals with maintainers
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • No positive for 4x loop unrolling on Ampere, so will keep 2x unrolling for Neoverse N1
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Performance on Altra is about 30%-40% lower than 8268.
      • Performance on Altra is slightly better than N1SDP.
      • IO testing is doable with specific PCIe slots, with which PCIe bandwidth is not the bottle-neck.
      • Try the VFs with DPDK plugin. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
    • Benchmark Altra vs Cascade 8268
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind


12/08/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Juraj Linkes
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
    • VPP Device
      • VPP device testing issues may be caused by XL710 i40e fw or kernel module.
        • Working with VPP/DPDK/Intel to root cause this issue. - Juraj
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
        • Which is acceptable by CSIT maitainers
      • LF will provide QSFP+ fiber switch for FD.io lab.
        • Vexxhost just has a spare one, and LF will buy it for FD.io lab, which will probably happen this month.
      • N1SDP shipment to FD.io
        • Govind will track the status
      • CSIT budget plan for 10G switch purchase in FD.io lab. - Juraj, Tina
        • Trishan de Lanerolle <tdelanerolle@linuxfoundation.org> from LF is working with Machiek to provide 10G switch.
        • Arm is required to present Arm achievement and plan to TSC.
          • Govind will prepare the slides
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • SOC id will be available on /proc entry starting from kernel version 5.9
        • Will investigate the details - Lijian
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Benchmarked cross-connect and TX queue is dropping packets
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • To optimize ethernet-input and avf-input node with NEON intrinsics
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 two proposals upstreamed
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch.
      • Have to repeat the testing in the future.
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Ampere altra server has some PCIe bugs.
      • Try the VFs with DPDK plugin. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
    • IPsec on Arm platform. - Govind
      • Apply prefetches and loop-unrolling with ipsec-out node
        • Didn't observe much performance improvement (2%) so far
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

12/1/2020

  • Attendees
    • Govindarajan Mohandoss
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
    • VPP Device
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
      • LF will provide QSFP+ fiber switch for FD.io lab.
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • To enable voting right for the VPP device jobs. - Juraj
          • Failed tests due to sw_interface_dump api issue. - Juraj
        • VPP device job is unstable
          • Race condition occurs when multiple VPP instances are starting.
          • Will try to update the i40e driver & firmware.
      • N1SDP shipment to FD.io
        • Govind will update the shippment status to Juraj and Machiek.
        • Will still have to ship N1SDP to FD.io lab, and Machiek confirmed that there will be enough rack space.
      • CSIT budget plan for 10G switch purchase in FD.io lab. - Juraj, Tina
        • Trishan de Lanerolle <tdelanerolle@linuxfoundation.org> from LF is working with Machiek to provide 10G switch.
        • Arm is required to present Arm achievement and plan to TSC.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • SOC id will be available on /proc entry starting from kernel version 5.9
        • Will investigate the details - Lijian
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • To optimize ethernet-input and avf-input node with NEON intrinsics
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 proposal
      • Refactored ethernet-input node with SVE/SVE2 intrinsics per Damjan's suggestion
      • Patches are upstreamed for comments
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch. - Jieqiang
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Ampere altra server has some PCIe bugs.
      • Try the VFs with DPDK plugin. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
    • IPsec on Arm platform. - Govind
      • Apply prefetches with ipsec-out node
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

11/24/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
    • VPP Device
      • Current VPP device testing on TX2 is around 40 mins - 45 mins
      • LF will provide QSFP+ fiber switch for FD.io lab.
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • To enable voting right for the VPP device jobs. - Juraj
          • Failed tests due to sw_interface_dump api issue. - Juraj
      • N1SDP shipment to FD.io
        • Govind will update the shippment status to Juraj and Machiek.
        • Will still have to ship N1SDP to FD.io lab, and Machiek confirmed that there will be enough rack space.
      • CSIT budget plan for 10G switch purchase in FD.io lab. - Juraj, Tina
        • Trishan de Lanerolle <tdelanerolle@linuxfoundation.org> from LF is working with Machiek to provide 10G switch.
        • Arm is required to present Arm achievement and plan to TSC.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • SOC id will be available on /proc entry starting from kernel version 5.9
        • Will investigate the details - Lijian
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • To optimize ethernet-input and avf-input node with NEON intrinsics
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 proposal
      • Refactored ethernet-input node with SVE/SVE2 intrinsics per Damjan's suggestion
      • Patches are upstreamed for comments
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch. - Jieqiang
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Ampere altra server has some PCIe bugs.
      • Try the VFs with DPDK plugin. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
    • IPsec on Arm platform. - Govind
      • Apply prefetches with ipsec-out node
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

11/17/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Tina Tsou
  • General
  • CSIT
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • SOC id will be available on /proc entry starting from kernel version 5.9
        • Will investigate the details - Lijian
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Initial benchmarking and analysis is done, and profiling result is recorded.
      • To optimize ethernet-input and avf-input node with NEON intrinsics
      • Try cross-connect with AVF PMD driver to check avf-input node only - Lijian
    • SVE/SVE2 proposal
      • Refactored ethernet-input node with SVE/SVE2 intrinsics per Damjan's suggestion
      • Patches are upstreamed for comments
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch. - Jieqiang
    • Benchmark the performance using AVF PMD driver on Ampere server. - Jieqiang
      • Ampere altra server has some PCIe bugs.
      • Try the VFs with DPDK plugin. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
    • IPsec on Arm platform. - Govind
      • Apply prefetches with ipsec-out node
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind

11/10/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
      • CSIT DPDK test cases will be enabled on Arm servers and data will be included next CSIT release report.
      • L2 learning 1Mx flows, 4T4C, with release-2005, about 20% performance drop.
        • The patch caused this issue has been identified - https://gerrit.fd.io/r/c/vpp/+/26549
        • Repeat tests on local N1SDP and cascade server. - Jieqiang
        • Repeat the test case with latest master branch. - Jieqiang
        • The patch introduced this perf drop need to be analyzed. - Jieqiang, Lijian
        • This patch needs to be analysed on VPP 2005 and 2001 releases. - Jieqiang, Lijian
        • The perf drop rate is ~5-8% on latest VPP code compared to the original data.
      • https://jenkins.fd.io/view/csit/job/csit-vpp-perf-mrr-daily-master-3n-tsh/
      • 1 Spare TX2 server can be used to create 2 node topology to run performance tests.
        • Juraj to check with Peter about the feasibility.
        • Move the thx2 to the same rack for tg and install the same nic on tg.
        • 1g NIC for management installed on thx2, but cannot be net-booted.
          • Able to net-boot from the built-in 10G NIC.
          • The tx2 has been moved to the same rack where the tg is located.
          • Plan to set up the weekly perf tests on the new topo.
        • Port the robotframe configuration steps for tsh testbeds from thx1 to thx2 to speed up perf tests. - Juraj
    • VPP Path
      • Totally 6x ThunderX1 servers in Nomad cluster
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Check with Dave about what we should do with CentOS-7 on Arm Jenkins if CentOS-8 is the main distro for VPP verification. - Juraj
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - First step;
        • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge' (Ubuntu-20.04 job will also be triggered)
        • https://gerrit.fd.io/r/c/ci-management/+/28960
        • https://gerrit.fd.io/r/c/ci-management/+/28022 automate the generation of docker builder images.
        • Plan to drop the support for CentOS 7 from Dave.
        • Tried Dave's patch to generate docker image on Arm and saw some errors. - Juraj
        • Test arm centos7 jenkins builder image. - Juraj.
        • Check with CSIT maintainers about the concrete plan for the enablement of Centos8 gerrit Jenkins job on aarch64. - Juraj
    • VPP Device
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • The servers are physically installed. Packages are installed. CSIT tests are run on these servers outside of jenkins. pxeboot (N/W boot) works fine with 10G NIC (Inband) and not with 1G NIC. One of the server works after reboot and the other server loses N/W connectivity.
        • Figure out which host of two hosts to run the Jenkins job.
        • Revert to old kernel version 4.15.0-55 to avoid AVF issue.
        • AVF issue is common across the platform.
          • Differences between avf driver versions may be the root cause of behavior changes.
        • New VPP device job takes about 55 minutes to finish, which needs to be reduced to 45 minutes around.
          • Python runs slower on new thx2 servers than 1-node skylake.
          • Try new version of Python(such as 3.8) or split the device tests into two parts.
          • Check how many CPUs get utilized for robot framework execution on thx2 server.
          • Two thunderx2 are running fine right now and the VPP device jobs are almost done.
          • Disabling hyperthreading on new thx2 will speed up the VPP device tests.
          • Enable the voting right for the VPP device jobs. - Juraj
            • Failed tests due to sw_interface_dump api issue. - Juraj
      • N1SDP shippment to FD.io
        • Get response from Maciek about the rack space and traffic generator availability.
      • CSIT budget plan for 10g switch purchase in FD.io lab. - Juraj, Tina
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • Nitin requires VPP generic image supporting 64B and 128B cache line CPU optimally at the same time - cannot be satisfied so far.
      • Detect the CPU type from firmware for Perseus-CPU servers, need to confirm with customers. - Lijian
      • Summarize the meeting minutes and action items. - Lijian
      • SOC id will be available on /proc entry from kernel version 5.9 - Ljian, Honnappa
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
      • Start benchmarking AVF PMD driver in VPP on N1SDP.
      • Investigate the performance gap using AVF PMD driver between N1SDP and Cascade Lake. - Lijian
      • avf_input and avf_output nodes don't comsume lots of CPU cycles than dpdk-related nodes do.
    • SVE/SVE2 proposal
      • Will send email to Damjan asking him to review
      • SVE proposal patch is upstreamed, call for comments - https://gerrit.fd.io/r/c/vpp/+/28986
      • No further comments from VPP community.
      • Apply the SVE/SVE2 intrinsics on refactoring ethernet-input node. - Lijian
        • SVE/SVE2 functionality to be tested on the new development platform.
        • Verify SVE/SVE2 code changes on simulator.
        • Try to run standalone SVE codes on the new FPGA platform.
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch. - Jieqiang
    • Benchmark the performance of L2/L3/ACL tests using AVF PMD driver on Ampere server. - Jieqiang
      • Ampere altra server has some PCIe bugs.
      • Try the VFs with DPDK plugin. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
      • Find out the tuned configuration for cross connect test cases using AVF PMD driver.
        • Figure out corresponding configurations in CSIT scripts.
        • Repeat the ACL ingress SL test cases locally for N1SDP.
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
        • Will investigate the prefetches and loop unrolling on IPsec input node. - Govind
    • Plans

11/03/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Totally 6x ThunderX1 servers in Nomad cluster
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Check with Dave about what we should do with CentOS-7 on Arm Jenkins if CentOS-8 is the main distro for VPP verification. - Juraj
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - First step;
        • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge' (Ubuntu-20.04 job will also be triggered)
        • https://gerrit.fd.io/r/c/ci-management/+/28960
        • https://gerrit.fd.io/r/c/ci-management/+/28022 automate the generation of docker builder images.
        • Test arm centos7 jenkins builder image. - Juraj.
        • Check with CSIT maintainers about the concrete plan for the enablement of Centos8 gerrit Jenkins job on aarch64. - Juraj
    • VPP Device
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • The servers are physically installed. Packages are installed. CSIT tests are run on these servers outside of jenkins. pxeboot (N/W boot) works fine with 10G NIC (Inband) and not with 1G NIC. One of the server works after reboot and the other server loses N/W connectivity.
        • Figure out which host of two hosts to run the Jenkins job.
        • Revert to old kernel version 4.15.0-55 to avoid AVF issue.
        • AVF issue is common across the platform.
          • Differences between avf driver versions may be the root cause of behavior changes.
        • New VPP device job takes about 55 minutes to finish, which needs to be reduced to 45 minutes around.
          • Python runs slower on new thx2 servers than 1-node skylake.
          • Try new version of Python(such as 3.8) or split the device tests into two parts.
          • Check how many CPUs get utilized for robot framework execution on thx2 server.
          • Two thunderx2 are running fine right now and the VPP device jobs are almost done.
      • N1SDP shippment to FD.io
        • Get response from Machiek about the rack space and traffic generator avalability.
      • CSIT budget plan for 10g switch purchase in FD.io lab. - Juraj, Tina
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • Nitin requires VPP generic image supporting 64B and 128B cache line CPU optimally at the same time - cannot be satisfied so far.
      • Detect the CPU type from firmware for Perseus-CPU servers, need to confirm with customers. - Lijian
      • Summarize the meeting minutes and action items. - Lijian
      • SOC id will be available on /proc entry from kernel version 5.9 - Ljian, Honnappa
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
      • Start benchmarking AVF PMD driver in VPP on N1SDP.
      • Investigate the performance gap using AVF PMD driver between N1SDP and Cascade Lake. - Lijian
      • avf_input and avf_output nodes don't comsume lots of CPU cycles than dpdk-related nodes do.
    • SVE/SVE2 proposal
      • Will send email to Damjan asking him to review
      • SVE proposal patch is upstreamed, call for comments - https://gerrit.fd.io/r/c/vpp/+/28986
      • No further comments from VPP community.
      • Apply the SVE/SVE2 intrinsics on refractoring ethernet-input node. - Lijian
        • SVE/SVE2 functionality to be tested on the new development platform.
    • Repeat the 4x and 2x loop unrolling tests on Ampere server for L3 forwarding with Lijian's patch. - Jieqiang
    • Benchmark the performance of L2/L3/ACL tests using AVF PMD driver on Ampere server. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
      • Find out the tuned configuration for cross connect test cases using AVF PMD driver.
        • Figure out corresponding configurations in CSIT scripts
        • Repeat the ACL ingress SL test cases locally for N1SDP.
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
        • Review akshitha's PPT on SLC eviction and share it with the team. - Govind.
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
    • Plans

10/27/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Totally 6x ThunderX1 servers in Nomad cluster
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Check with Dave about what we should do with CentOS-7 on Arm Jenkins if CentOS-8 is the main distro for VPP verification. - Juraj
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - First step;
        • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge' (Ubuntu-20.04 job will also be triggered)
        • https://gerrit.fd.io/r/c/ci-management/+/28960
        • Check with CSIT maintainers about the concrete plan for the enablement of Centos8 gerrit Jenkins job on aarch64. - Juraj
    • VPP Device
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • The servers are physically installed. Packages are installed. CSIT tests are run on these servers outside of jenkins. pxeboot (N/W boot) works fine with 10G NIC (Inband) and not with 1G NIC. One of the server works after reboot and the other server loses N/W connectivity.
        • Figure out which host of two hosts to run the Jenkins job.
        • Revert to old kernel version 4.15.0-55 to aviod AVF issue.
          • Differences between avf driver versions may be the root cause of behavior changes.
        • New VPP device job takes about 55 minutes to finish, which needs to be reduced to 40 minutes around.
          • Python runs slower on new thx2 servers than 1-node skylake.
          • Try new version of Python(such as 3.8) or split the device tests into two parts.
          • Check how many CPUs get utilized for robot framework execution on thx2 server.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • Nitin requires VPP generic image supporting 64B and 128B cache line CPU optimally at the same time - cannot be satisfied so far.
      • Detect the CPU type from firmware for Perseus-CPU servers, need to confirm with customers. - Lijian
      • Summarize the meeting minutes and action items. - Lijian
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
      • Start benchmarking AVF PMD driver in VPP on N1SDP.
      • Investigate the performance gap using AVF PMD driver between N1SDP and Cascade Lake. - Lijian
      • avf_input and avf_output nodes don't comsume lots of CPU cycles than dpdk-related nodes do.
    • SVE/SVE2 proposal
      • Will send email to Damjan asking him to review
      • SVE proposal patch is upstreamed, call for comments - https://gerrit.fd.io/r/c/vpp/+/28986
      • No further comments from VPP community.
      • Apply the SVE/SVE2 on ethernet-input node. - Lijian
    • Repeat the 4x and 2x loop unrolling tests on Ampere server. - Jieqiang
    • Benchmark the performance of L2/L3/ACL tests using AVF PMD driver on Ampere server. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
        • Figure out corresponding configurations in CSIT scripts
        • Repeat the ACL ingress SL test cases locally for N1SDP.
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
      • Work on IPsec input node and VPP uses linear search on SPD lookup.
        • Will try loop unrolling on the SPD lookup.
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
        • Tuned version has higher SLC eviction than the default version, talk with CPU team to figure out the reason.
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
    • Plans

10/20/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Totally 6x ThunderX1 servers in Nomad cluster
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Check with Dave about what we should do with CentOS-7 on Arm Jenkins if CentOS-8 is the main distro for VPP verification. - Juraj
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - First step;
        • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge' (Ubuntu-20.04 job will also be triggered)
        • https://gerrit.fd.io/r/c/ci-management/+/28960
        • Check with CSIT maintainers about the concrete plan for the enablement of Centos8 gerrit Jenkins job on aarch64. - Juraj
        • Errors happen when running latest VPP debug image, which was introduced by https://gerrit.fd.io/r/c/vpp/+/29490 - Lijian
    • VPP Device
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • The servers are physically installed. Packages are installed. CSIT tests are run on these servers outside of jenkins. pxeboot (N/W boot) works fine with 10G NIC (Inband) and not with 1G NIC. One of the server works after reboot and the other server loses N/W connectivity.
        • Figure out which host of two hosts to run the Jenkins job.
        • Two failed test cases related to AVF plugin.
          • The root cause is the newer kernel version - 4.15.0-118-generic fails, 4.15.0-72-generic works.
          • Downgrade the kernel version to 4.15.0-72-generic and continue the VPP device testing.
          • Try the same experiment on X86 to see if this issue is arm-specific or not. - Juraj
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • Nitin requires VPP generic image supporting 64B and 128B cache line CPU optimally at the same time - cannot be satisfied so far.
      • Detect the CPU type from firmware for Perseus-CPU servers, need to confirm with customers. - Lijian
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
      • Start benchmarking AVF PMD driver in VPP on N1SDP.
      • Investigate the performance gap using AVF PMD driver between N1SDP and Cascade Lake. - Lijian
    • SVE/SVE2 proposal
    • Benchmark VPP scalability on N1SDP vs CascadeLake, with 3x CPUs.
      • Finished the benchmarking and shared the data to team. - Jieqiang
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
        • Figure out corresponding configurations in CSIT scripts
        • Repeat the ACL ingress SL test cases locally for N1SDP.
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
      • Tune PMD config to achieve zero cache eviction for System Level Cache on N1SDP - kamalakshitha
      • Prefetches on IPsec and learn about the cache behavior. - kamalakshitha
    • Plans

10/13/2020

  • Attendees
    • Govindarajan Mohandoss
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
    • Tina Tsou
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Totally 6x ThunderX1 servers in Nomad cluster
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - First step;
        • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge' (Ubuntu-20.04 job will also be triggered)
        • https://gerrit.fd.io/r/c/ci-management/+/28960
        • Check with CSIT maintainers about the concrete plan for the enablement of Centos8 gerrit Jenkins job on aarch64. - Juraj
    • VPP Device
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • The servers are physically installed. Packages are installed. CSIT tests are run on these servers outside of jenkins. pxeboot (N/W boot) works fine with 10G NIC (Inband) and not with 1G NIC. One of the server works after reboot and the other server loses N/W connectivity.
        • Figure out which host of two hosts to run the Jenkins job.
        • Two failed test cases related to AVF plugin.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs with other Perseus CPUs
      • Nitin requires VPP generic image supporting 64B and 128B cache line CPU optimally at the same time - cannot be satisfied so far.
      • Detect the CPU type from firmware for Perseus-CPU servers, need to confirm with customers. - Lijian
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
    • SVE/SVE2 proposal
    • Benchmark VPP scalability on N1SDP vs CascadeLake, with 3x CPUs.
      • Finished the benchmarking and shared the data to team.
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
        • Figure out corresponding configurations in CSIT scripts
        • Repeat the ACL ingress SL test cases locally for N1SDP.
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
    • Plans

10/06/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Tina Tsou
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Totally 6x ThunderX1 servers in Nomad cluster
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - First step;
        • CentOS-8 on Arm Jenkins is created and could be triggered manually with 'beta-verify' and 'beta-merge' (Ubuntu-20.04 job will also be triggered)
        • https://gerrit.fd.io/r/c/ci-management/+/28960
    • VPP Device
      • CSIT will install normally used os distro and kernel.
      • 2x ThunderX2 servers are setup in FD.io lab. SSH and IPMI connections are working.
        • The servers are physically installed. Packages are installed. CSIT tests are run on these servers outside of jenkins. pxeboot (N/W boot) works fine with 10G NIC (Inband) and not with 1G NIC. One of the server works after reboot and the other server loses N/W connectivity.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate vendor CPUs and other Perseus CPUs
      • Nitin requires VPP generic image supporting 64B and 128B cache line CPU optimally at the same time - cannot be satisfied so far.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
    • SVE/SVE2 proposal
    • Benchmark VPP scalability on N1SDP vs CascadeLake, with 3x CPUs.
      • Finished the benchmarking and shared the data to team.
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
      • Figure out corresponding configurations in CSIT scripts
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
    • Plans

09/29/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate Vendor CPUs with other Perseus CPUs
      • Nitin requires VPP generic image supporting 64B and 128B cache line CPU optimally at the same time - cannot be satisfied so far.
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
    • SVE/SVE2 proposal
    • Benchmark VPP scalability on N1SDP vs CascadeLake, with 3x CPUs.
      • Finished the benchmarking and shared the data to team.
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
      • Figure out corresponding configurations in CSIT scripts
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
    • Plans

09/22/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Multi-arch support - Lijian
      • Key point is how to differentiate Vendor CPUs with other Perseus CPUs
    • Investigate VPP Intel AVF PMD driver - Lijian
      • Start investigating AVF code in VPP.
    • SVE/SVE2 proposal
    • Benchmark VPP scalability on N1SDP vs CascadeLake, with 3x CPUs.
      • Repeat scalability testing on N1SDP, e.g., 1T1C, 2T2C, 3T3C
      • Figure out corresponding configurations in CSIT scripts
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • IPsec on Arm platform. - Govind
    • Plans

09/15/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • LF will pay for the expense, and Vexhost has or will make the order for new RAM module.
      • Will confirm with Dean if Arm can pay for the expense. If yes, will send the proposal to vexhost.
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
      • Check with Juraj with the latest news about the faulty RAMs.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - first step;
        • Add CentOS-7 on Arm will be second step.
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
        • Juraj/Jieqiang to help Dave Wallace to fix the script issues. Currently, the build process is done manually and will be automated.
    • VPP Device
      • We can decommission 3x SoftIron servers directly, but for the existing ThunderX2 servers, the decommission with it could be temporarily. We probably will reinstall it in the near future.
      • Mention the rack space request for the two ThunderX2 servers in CSIT meeting. - Juraj
    • Two ThunderX2 servers are received by Vexx host and currently in the storage warehouse.
    • Vexx host people will setup the servers and provide IP connectivity. Juraj will install the necessary software after that.
    • Budget plan for CSIT FD.io lab.
      • We have enough servers for VPP path & device tests.
      • We can ask the CSIT FD.io lab folks for saving rack space for arm servers.
      • We may plan to send new advanced servers for perf tests in future but we won't mention the specific server type.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • Vendor CPU server enablement in VPP - Lijian
      • Ready for internal review
      • Will discuss with VPP maintainer
    • Investigate VPP Intel AVF driver - Lijian
    • SVE
      • SVE intrinsics wrapper is done. Proposal patch is ready for review.
      • ACLE, architecture, sve-sve2-programming-example
      • SVE intrinsics is preferred.
      • Share dpdk team with SVE knowledge.
    • Benchmarked VPP on n1sdp on scalability, on 3x CPUs.
      • Will repeat scalability testing on N1SDP.
    • Benchmark AVF driver btw Cascade Lake and N1SDP - Jieqiang
      • Will investigate AVF drivers on Arm. - Lijian
    • Jieqiang will figure out performance data for 1x, 10Kx flows on Cascade Lake in CSIT.
      • Conform if the system is same for the local dell server and cascade server in CSIT. - Jieqiang
      • Check if there are any test cases with 1t1c/2t2c/4t4c configured for 2n-clx testbed in CSIT - Jieqiang
      • Performance data; Configurations;
    • VM2VM
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Started system tuning on PMD TX direction.
      • Investigate mempool configuration.
      • Change the descriptor size by modifying the DPDK source code.
      • Currently working on non-encryption optimization with PMD driver.
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans

09/08/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • LF will pay for the expense, and Vexhost has or will make the order for new RAM module.
      • Will confirm with Dean if Arm can pay for the expense. If yes, will send the proposal to vexhost.
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Dave is preparing scripts to generate docker images automatically on both x86 and Arm - first step;
        • Add CentOS-7 on Arm will be second step.
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
        • Juraj/Jieqiang to help Dave Wallace to fix the script issues. Currently, the build process is done manually and will be automated.
    • VPP Device
      • We can decommission 3x SoftIron servers directly, but for the existing ThunderX2 servers, the decommission with it could be temporarily. We probably will reinstall it in the near future.
      • Mention the rack space request for the two ThunderX2 servers in CSIT meeting. - Juraj
    • Two ThunderX2 servers are received by Vexx host and currently in the storage warehouse.
    • Vexx host people will setup the servers and provide IP connectivity. Juraj will install the necessary software after that.
  • VPP
    • In Ubuntu-20.04, VPP on Arm will reprioritize compiler as gcc-10 > gcc-9.2.0 > clang-10
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • SVE
      • SVE intrinsics wrapper is done. Proposal patch is ready for review.
      • ACLE, architecture, sve-sve2-programming-example
      • SVE intrinsics is preferred.
    • Benchmarked VPP on n1sdp on scalability, on 3x CPUs.
      • Will repeat scalability testing on N1SDP.
    • Benchmark AVF driver btw Cascade Lake and N1SDP - Jieqiang
      • Will investigate AVF drivers on Arm. - Lijian
    • Jieqiang will figure out performance data for 1x, 10Kx flows on Cascade Lake in CSIT.
      • Performance data; Configurations;
    • VM2VM
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Started system tuning on PMD TX direction.
      • Currently working on non-encryption optimization with PMD driver.
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans

09/01/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Will confirm with Dean if Arm can pay for the expense. If yes, will send the proposal to vexhost.
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
        • Seems plugin working RAMs into empty slots will resolve the problem.
        • Juraj will send email to Machiek about the ownership of any FD.io lab servers, and who should pay for the charge.
      • The second ThunderX1 has IPMI problem, but SSH is working fine.
        • IPMI IP is configured via SSH Linux prompt. It's working fine now.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • 3x spare ThunderX servers are used for CI and included in Nomad cluster.
        • Faulty RAM on TX server is not fixed and yet to be debugged.
        • Pending with Vexx host to proceed further.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
        • Juraj/Jieqiang to help Dave Wallace to fix the script issues. Currently, the build process is done manually and will be automated.
    • VPP Device
      • Mention the rack space request for the two ThunderX2 servers in CSIT meeting. - Juraj
    • Two ThunderX2 servers are received by Vexx host and currently in the storage warehouse.
    • Vexx host people will setup the servers and provide IP connectivity. Juraj will install the necessary software after that.
  • VPP
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
      • gcc-10.1.0 has compiling errors with latest VPP source code.
        • This issue is fixed by Jieqiang and available for internal review.
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • gcc-10 compiling issue is resolved and merged.
    • SVE
      • ACLE, architecture, sve-sve2-programming-example
      • SVE intrinsics is preferred.
    • Benchmarked VPP on n1sdp on scalability, on 3x CPUs.
    • VM2VM
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Started system tuning on PMD TX direction.
      • Currently working on non-encryption optimization with PMD driver.
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans

08/25/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
        • Seems plugin working RAMs into empty slots will resolve the problem.
        • Juraj will send email to Machiek about the ownership of any FD.io lab servers, and who should pay for the charge.
      • The second ThunderX1 has IPMI problem, but SSH is working fine.
        • IPMI IP is configured via SSH Linux prompt. It's working fine now.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • 3x spare ThunderX servers are used for CI and included in Nomad cluster.
        • Faulty RAM on TX server is not fixed and yet to be debugged.
        • Pending with Vexx host to proceed further.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
        • Juraj/Jieqiang to help Dave Wallace to fix the script issues. Currently, the build process is done manually and will be automated.
    • VPP Device
      • Mention the rack space request for the two ThunderX2 servers in CSIT meeting. - Juraj
    • Two ThunderX2 servers are received by Vexx host and currently in the storage warehouse.
    • Vexx host people will setup the servers and provide IP connectivity. Juraj will install the necessary software after that.
  • VPP
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
      • gcc-10.1.0 has compiling errors with latest VPP source code.
        • This issue is fixed by Jieqiang and available for internal review.
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • SVE
      • ACLE, architecture, sve-sve2-programming-example
      • SVE intrinsics is preferred.
    • Benchmarked VPP on n1sdp on scalability, on 3x CPUs.
    • VM2VM
    • Transport use cases on VPP. - Govind
      • Discussed the node graph and topology.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Currently working on non-encryption optimization with PMD driver.
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans


08/18/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
      • Jieqiang is investigating some performance drop (between 2005 and 2008 releases) cases on Taishan servers.
      • https://jenkins.fd.io/view/csit/job/csit-vpp-perf-mrr-daily-master-3n-tsh/
      • They have finished collecting data with performance testing setup, and the mrr daily is resumed
      • FD.io CSIT-2005 Release Report was released, https://docs.fd.io/csit/rls2005/report/
      • Jieqiang will share investigation report, but so far there is no apparent performance differences.
        • Some 4T4C test cases on Taishan have obvious performance drop. Will compare the trending with x86 machines.
    • VPP Path
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
      • The second ThunderX1 has IPMI problem, but SSH is working fine.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • 3x spare ThunderX servers are used for CI and included in Nomad cluster.
        • Faulty RAM on TX server is not fixed and yet to be debugged.
        • Pending with Vexx host to proceed further.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
        • Juraj/Jieqiang to help Dave Wallace to fix the script issues. Currently, the build process is done manually and will be automated.
    • VPP Device
    • Two ThunderX2 servers are received by Vexx host and currently in the storage warehouse.
    • Vexx host people will setup the servers and provide IP connectivity. Juraj will install the necessary software after that.
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
      • Initial benchmarking with VPP hostack on N1SDP was done. 29Gb/s on N1SDP and 26Gb/s on Haswell.
      • Investigating vlib_timer and timer wheel in VPP.
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
      • gcc-10.1.0 has compiling errors with latest VPP source code.
        • This issue is fixed by Jieqiang and available for internal review.
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Currently working on non-encryption optimization with PMD driver.
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans


08/11/2020

  • Attendees
    • Honnappa Nagarahalli
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
    • Filip Varga
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
      • The second ThunderX1 has IPMI problem, but SSH is working fine.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • 3x spare ThunderX servers are used for CI and included in Nomad cluster.
        • Faulty RAM on TX server is not fixed and yet to be debugged.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • Two ThunderX2 servers have been collected and shipped, and will target to arrive on July 30th.
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
      • Initial benchmarking with VPP hostack on N1SDP was done. 29Gb/s on N1SDP and 26Gb/s on Haswell.
      • Investigating vlib_timer and timer wheel in VPP.
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
      • gcc-10.1.0 has compiling errors with latest VPP source code.
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Currently working on non-encryption optimization with PMD driver.
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans

08/04/2020

  • Attendees
    • Honnappa Nagarahalli
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
    • Filip Varga
  • General
  • CSIT
    • VPP Performance Test
      • They have finished collecting data with performance testing setup, and the mrr daily is resumed
      • FD.io CSIT-2005 Release Report was released, https://docs.fd.io/csit/rls2005/report/
      • Jieqiang will share investigation report, but so far there is no apparent performance differences.
        • Some 4T4C test cases on Taishan have obvious performance drop. Will compare the trending with x86 machines.
    • VPP Path
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
      • The second ThunderX1 has IPMI problem, but SSH is working fine.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • 3x spare ThunderX servers are used for CI and included in Nomad cluster.
        • Faulty RAM on TX server is not fixed and yet to be debugged.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • Two ThunderX2 servers have been collected and shipped, and will target to arrive on July 30th.
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
      • Initial benchmarking with VPP hostack on N1SDP was done. 29Gb/s on N1SDP and 26Gb/s on Haswell.
      • Investigating vlib_timer and timer wheel in VPP.
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
      • gcc-10.1.0 has compiling errors with latest VPP source code.
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Currently working on non-encryption optimization with PMD driver.
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans


07/28/2020

  • Attendees
    • Honnappa Nagarahalli
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
    • VPP Performance Test
      • They have finished collecting data with performance testing setup, and the mrr daily is resumed
      • FD.io CSIT-2005 Release Report was released, https://docs.fd.io/csit/rls2005/report/
      • Jieqiang will share investigation report, but so far there is no apparent performance differences.
      • VPP performance testing is running once a week.
    • VPP Path
      • One ThunderX1 has faulty RAM. Will try to replace all the RAMs.
      • The second ThunderX1 has IPMI problem, but SSH is working fine.
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • 3x spare ThunderX servers are used for CI and included in Nomad cluster.
        • Faulty RAM on TX server is not fixed and yet to be debugged.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Will confirm with Dave W. if he will add this Jenkins job and if he requires any help - Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been upstreamed for review and merge.
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • Two ThunderX2 servers have been collected and shipped, and will target to arrive on July 30th.
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
      • Initial benchmarking with VPP hostack on N1SDP was done. 29Gb/s on N1SDP and 26Gb/s on Haswell.
      • Investigating vlib_timer and timer wheel in VPP.
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
    • N1SDP enablement. - Lijian
      • Preparing patches to enable creating big tables on huge-pages
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Focus both non-encryption and encryption cases.
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans

07/21/2020

  • Attendees
    • Honnappa Nagarahalli
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
    • VPP Performance Test
    • VPP Path
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • 3x spare ThunderX servers are used for CI and included in Nomad cluster. 1 Debugging server for VPP Dev and 3 servers (2 TX and 1 TX2) are unreachable through IPMI and one of them is reachable through SSH. IPMI unreachability is still investigated by Vexx host. CI functionality is restored with spare TX servers. TX2 server is unreachable through IPMI and VPP device jobs are not running. Faulty RAM on TX server is not fixed and yet to be debugged.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • Arm has
    • If vexxhost can collect the hardware, will ship the servers asap.
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
      • Initial benchmarking with VPP hostack on N1SDP was done. 29Gb/s on N1SDP and 26Gb/s on Haswell.
      • Investigating vlib_timer and timer wheel in VPP.
    • Benchmarking btw gcc-10.1.0, clang-10/clang-9 and gcc-9.2.0 on Arm machines - Jieqiang
    • N1SDP enablement. - Lijian
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
        • Upstreamed and are using csit testing to verify the patch.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans


07/14/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
      • VPP performance testing is running once a week.
      • Community has started collecting performance data with these CSIT machines.
    • VPP Path
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • Will probably use 3xspare ThunderX1 servers as CI build server/nomad cluster.
        • Two of the three ThunderX1 servers cannot be accessed.
        • Spare ThunderX servers are used for CI and included in Nomad cluster. 1 Debugging server for VPP Dev and 3 servers (2 TX and 1 TX2) are unreachable through IPMI and one of them is reachable through SSH. IPMI unreachability is still investigated by Vexx host. CI functionality is restored with spare TX servers. TX2 server is unreachable through IPMI and VPP device jobs are not running. Faulty RAM on TX server is not fixed and yet to be debugged.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been verified by Jieqiang, will send to Dave Wallace to use it for VPP Jenkins job.
        • Jieqiang will send email to Dave Wallace about CentOS-7 on Arm Jenkins job. - Jieqiang
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • If vexxhost can collect the hardware, will ship the servers asap.
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
      • 'make build-release CC=gcc' will override default clang-9 in vpp.
    • N1SDP enablement. - Lijian
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
      • Investigating various No. of rx_q_bufs & tx_q_bufs
      • Investigating various No. of vector size, and check its affection on throughput
      • Benchmark and compare PMU counters btw 4x and 2x loop unrolling on n1sdp
    • ACL optimization investigation on n1sdp - Govind
      • Investigating using SPE counters to profile ACL plugin bottle-neck
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

07/07/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
      • VPP performance testing is running once a week.
      • Community has started collecting performance data with these CSIT machines.
    • VPP Path
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • Will probably use 3xspare ThunderX1 servers as CI build server/nomad cluster.
        • Two of the three ThunderX1 servers cannot be accessed.
        • Spare ThunderX servers are used for CI and included in Nomad cluster. 1 Debugging server for VPP Dev and 3 servers (2 TX and 1 TX2) are unreachable through IPMI and one of them is reachable through SSH. IPMI unreachability is still investigated by Vexx host. CI functionality is restored with spare TX servers. TX2 server is unreachable through IPMI and VPP device jobs are not running. Faulty RAM on TX server is not fixed and yet to be debugged.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been verified by Jieqiang, will send to Dave Wallace to use it for VPP Jenkins job.
        • Jieqiang will send email to Dave Wallace about CentOS-7 on Arm Jenkins job. - Jieqiang
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • If vexxhost can collect the hardware, will ship the servers asap.
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
      • 'make build-release CC=gcc' will override default clang-9 in vpp.
    • N1SDP enablement. - Lijian
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
      • Investigating various No. of rx_q_bufs & tx_q_bufs
      • Investigating various No. of vector size, and check its affection on throughput
      • Benchmark and compare PMU counters btw 4x and 2x loop unrolling on n1sdp
    • ACL optimization investigation on n1sdp - Govind
      • Investigating using SPE counters to profile ACL plugin bottle-neck
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang


06/30/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
      • VPP performance testing is running once a week.
      • Community has started collecting performance data with these CSIT machines.
    • VPP Path
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • Will probably use 3xspare ThunderX1 servers as CI build server/nomad cluster.
        • Two of the three ThunderX1 servers cannot be accessed.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been verified by Jieqiang, will send to Dave Wallace to use it for VPP Jenkins job.
        • Jieqiang will send email to Dave Wallace about CentOS-7 on Arm Jenkins job. - Jieqiang
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • If vexxhost can collect the hardware, will ship the servers asap.
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • VPP hoststack TCP/CPS(Connnection per Second) investigation;
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
      • 'make build-release CC=gcc' will override default clang-9 in vpp.
    • N1SDP enablement. - Lijian
      • IP4-rewrite refactor patch brings performance improvement especially with 10K flows
      • Investigating various No. of rx_q_bufs & tx_q_bufs
      • Investigating various No. of vector size, and check its affection on throughput
      • Benchmark and compare PMU counters btw 4x and 2x loop unrolling on n1sdp
    • ACL optimization investigation on n1sdp - Govind
      • Investigating using SPE counters to profile ACL plugin bottle-neck
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

06/23/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
    • VPP Performance Test
      • VPP performance testing is running once a week.
      • Community has started collecting performance data with these CSIT machines.
    • VPP Path
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
        • Two of the three ThunderX1 servers cannot be accessed.
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload.
        • The Dockerfile has been verified by Jieqiang, will send to Dave Wallace to use it for VPP Jenkins job.
        • Jieqiang will send email to Dave Wallace about CentOS-7 on Arm Jenkins job. - Jieqiang
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • If vexxhost can collect the hardware, will ship the servers asap.
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • L3FWD status
    • CSIT status
    • EPIC plan
      • SVE2 investigation in VPP;
      • VPP hoststack TCP/CPS(Connnection per Second) investigation;
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
      • 'make build-release CC=gcc' will override default clang-9 in vpp.
    • N1SDP enablement. - Lijian
      • Profiling with NMU-600 counters.
      • Investigate 4x loop unrolling performance degradation issue.
      • Throughput performance drop as flow number increases in N1SDP.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

06/16/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
    • VPP Performance Test
      • VPP performance testing is running once a week.
      • Community has started collecting performance data with these CSIT machines.
    • VPP Path
      • Juraj will follow or create new vexxhost ticket to replace faulty RAM.
      • Investigate Ubuntu-20.04 on Arm servers - Jieqiang
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload. The docker file needs to be tested with local VPP sand box before uploading. The docker file needs to be
        • labelled by Dave Wallace to use it for VPP Jenkins job.
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
        • gcc-9 is hard-coded and used, so compilation issue is gone.
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • If vexxhost can collect the hardware, will ship the servers asap.
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
      • 'make build-release CC=gcc' will override default clang-9 in vpp.
    • N1SDP enablement. - Lijian
      • Profiling with NMU-600 counters.
      • Investigate 4x loop unrolling performance degradation issue.
      • Throughput performance drop as flow number increases in N1SDP.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

06/09/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
    • VPP Performance Test
      • VPP performance testing is running once a week.
      • Community will collect performance data with these CSIT machines.
      • IPSec tunnel configuration issue.
        • Issue is resolved.
        • Also failing on x86. CSIT maintainer is trying to root cause the problem.
          • Juraj to run the IPSec regression on Taishan server with the IPSec patch.
    • VPP Path
      • Investigate Ubuntu-20.04 on Arm servers - Jieqiang
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload. The docker file needs to be tested with local VPP sand box before uploading. The docker file needs to be
        • labelled by Dave Wallace to use it for VPP Jenkins job.
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
        • gcc-9 is hard-coded and used, so compilation issue is gone.
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
      • Two ThunderX2 information will be confirmed with FD.io CSIT lab admin. - Jieqiang
      • Commit internal patch to support ThunderX2 in VPP device testing. - Jieqiang
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • If vexxhost can collect the hardware, will ship the servers asap.
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
      • 'make build-release CC=gcc' will override default clang-9 in vpp.
    • N1SDP enablement. - Lijian
      • Profiling with NMU-600 counters.
      • Investigate 4x loop unrolling performance degradation issue.
      • Throughput performance drop as flow number increases in N1SDP.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

06/02/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
    • Lijian Zhang
  • General
  • CSIT
    • VPP Performance Test
      • IPSec tunnel configuration issue.
        • Also failing on x86. CSIT maintainer is trying to root cause the problem.
          • Juraj to run the IPSec regression on Taishan server with the IPSec patch.
    • VPP Path
      • Investigate Ubuntu-20.04 on Arm servers - Jieqiang
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload. The docker file needs to be tested with local VPP sand box before uploading. The docker file needs to be
        • labelled by Dave wallace to use it for VPP Jenkins job.
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
        • gcc-9 is hard-coded and used, so compilation issue is gone.
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers.
        • Internal patch is committed. Requires legal permission.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
      • 'make build-release CC=gcc' will override default clang-9 in vpp.
    • N1SDP enablement. - Lijian
      • Investigate 4x loop unrolling performance degradation issue.
      • Throughput performance drop as flow number increases in N1SDP.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

05/26/2020

  • Attendees
    • Govindarajan Mohandoss
    • Juraj Linkes
    • Jieqiang Wang
    • Tina Tsou
  • General
  • CSIT
    • VPP Performance Test
      • IPSec tunnel configuration issue.
        • Also failing on x86. CSIT maintainer is trying to root cause the problem.
          • Juraj to run the IPSec regression on Taishan server with the IPSec patch.
    • VPP Path
      • Investigate Ubuntu-20.04 on Arm servers - Jieqiang
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
        • Questions on the docker file upload. The docker file needs to be tested with local VPP sand box before uploading. The docker file needs to be
        • labelled by Dave wallace to use it for VPP Jenkins job.
        • Vanessa Valderrama <vvalderrama@linuxfoundation.org>
        • 'Dave Wallace' <dwallacelf@gmail.com>
        • https://gerrit.fd.io/r/gitweb?p=ci-management.git;a=summary
        • gcc-9 is hard-coded and used, so compilation issue is gone.
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
    • Dean will schedule shipping these two TX2 servers to FD.io lab.
      • Update the document with server information before shipping the servers. Jieqiang will setup a meeting with Juraj regarding this documentation.
    • Update server information to CSIT documentation. - Juraj & Jieqiang
    • Prepare CSIT script patch for adding those two ThunderX2 servers. - Juraj & Jieqiang
      • This change can be done once TX2 servers are shipped to FDIO lab.
    • Dave wallace - Install nomad service in those two servers - Juraj & Jieqiang
      • Nomad takes care of redundancy and resources like CPU/Memory. 16 cores per job and 6 jobs in total.
    • The servers, intel NICs, and Mellanox NICs works good so far.
      • Root-causing the RDMA issue with Mellanox NIC.
    • ThunderX2 servers are in Arm local lab. Dean is setting up the hardware.
    • Two more ThunderX2 have just been ordered and are expected to arrive in Arm lab in April.
      • We are about to purchase two official ThunderX2 servers in market.
      • Raise the budget requirement from CE-OSS - Dean & Honnappa
      • Check the ThunderX2 configurations required - Govind & Juraj
    • Two ThunderX2 servers are installed in Arm lab.
    • Per patch regression: 2 node topology is freely available. New ARM setup can be made to run per patch regression.
  • VPP
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
    • N1SDP enablement. - Lijian
      • Investigate 4x loop unrolling performance degradation issue.
      • Throughput performance drop as flow number increases in N1SDP.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang


05/19/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
    • Juraj Linkes
    • Tina Tsou
    • Jieqiang Wang
    • Lijian Zhang
  • General
  • CSIT
    • VPP Performance Test
      • the other failure is related with VPP image on Arm, IPSec tunnel configuration issue.
        • Also failing on x86. CSIT maintainer is trying to root cause the problem.
    • VPP Path
      • Investigate Ubuntu-20.04 on Arm servers - Jieqiang
      • Investigate adding CentOS-7 on Arm Jenkins jobs - Juraj & Jieqiang
      • By fixing software bug, VPP can boot up normally with 16K/64K page size.
        • There's about 4-5 test failures in 'make test' when system is configured with 16K/64K page size - Lijian
    • VPP Device
    • VPP device job is running now and will be triggered per VPP patch and CSIT patch
  • FD.io lab
  • VPP
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
      • Will benchmark btw gcc-9 and clang-10 to decide which should be the default compiler, will sync up with Suresh.
    • N1SDP enablement. - Lijian
      • Investigate 4x loop unrolling performance degradation issue.
      • Throughput performance drop as flow number increases in N1SDP.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
      • Upstream the ACL patch for CSIT performance testing experiment.
    • Trying to make IPsec enabled with Arm platform. - Govind
      • Basic IPsec functions are working. Will do benchmarking per CPU core.
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

04/28/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
    • Juraj Linkes
    • Tina Tsou
    • Jieqiang Wang
  • General
  • CSIT
  • FD.io lab
  • VPP
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
      • Resolve vectorized endianness conversion error in Mellanox RDMA driver.
      • To benchmark Mellanox DPDK PMD driver vs RDMA driver in VPP. - Lijian
      • Confirm with Suresh on his benchmarking data/scripts with Mellanox NICs
    • Resolve VPP compiling issue with clang-6.
    • VPP default compiler is clang-9 now, which does not support optimized options -mcpu=neoverse-n1/-mtune=neoverse-n1
    • N1SDP enablement. - Lijian
      • Multi-arch, arch-specific compiling and dynamic function selection patch is merged.
      • IOMMU limitation issue is gone after upgrade the kernel and fw
        • Share kernel/fw upgrade version to Govind
      • Investigate 4x loop unrolling performance degradation issue.
      • Throughput performance drop as flow number increases in N1SDP.
    • ACL optimization investigation on n1sdp - Govind
      • Patch to remove redundancy prefetches are committed - Govind
      • Filed a confluence page to record the ACL investigation.
    • Trying to make IPsec enabled with Arm platform. - Govind
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • Will try to run VPP device testing on local ThunderX2 servers with XL710 25G NICs.
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang

04/28/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
    • Juraj Linkes
    • Tina Tsou
    • Jieqiang Wang
    • Arthur Marshall
  • General
  • CSIT
  • FD.io lab
  • VPP
    • Enabled RDMA driver on Arm, and Mellanox DPDK PMD driver is also working on ThunderX2
    • Investigate bihash operations in L2 throughput are hot-spots
      • To confirm with Damjan if he has plan to rewrite l2-nodes or not - Lijian
      • To confirm CRC32 calculation is compiled with O2 or O3 options - Lijian
    • Benchmarking, profiling and comparation btw VPP and testpmd are done. Will review with team. - Lijian
    • N1SDP enablement. - Lijian
    • GCC-9.2.0 is used with "-march=armv8.2-a+crc+crypto -mtune=neoverse-n1" compiler options.
      • Multi-arch, arch-specific compiling and dynamic function selection patch is ready.
      • It seems dual loop-unrolling gives better performance over quad-loop-unrolling.
    • iova_mode == VA not working issue is not root-caused
      • DMA mapping btw iova & pa; VPP and DPDK are using va as iova, and then do the DMA mapping.
      • However IOMMU on N1SDP requires a limited memory space, less than 40 bits?.
        • This issue will not be seen in the latest N1 Firmware. Upgrading to latest Firmware is pending.
    • Share L2/L3/ACL throughput wiz & wo L3 cache - Govind
    • Will try with L3 cache enabled to see if performance drop as flow number increasing issue is fixed or not. - Govind
      • The degradation is seen even when L3 cache is enabled.
    • Trying to make IPsec enabled with Arm platform. - Govind
    • Create Confluence page to record all the performance benchmarking data - Lijian
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang
        • Jieqiang needs 2 Intel NICs to make the test bed ready for VPP Path tests. Jieqiang and Lijian to discuss with Juraj to run the jenkins job on CentOS.

04/21/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
    • Juraj Linkes
    • Tina Tsou
    • Jieqiang Wang
  • General
  • CSIT
  • FD.io lab
  • VPP
    • Investigate bihash operations in L2 throughput are hot-spots
      • To confirm with Damjan if he has plan to rewrite l2-nodes or not - Lijian
      • To confirm CRC32 calculation is compiled with O2 or O3 options - Lijian
    • Benchmarking, profiling and comparation btw VPP and testpmd are done. Will review with team. - Lijian
    • N1SDP enablement. - Lijian
    • gcc-10 is not working so far.
      • GCC-9.2.0 is used with "-march=armv8.2-a+crc+crypto -mtune=neoverse-n1" compiler options.
      • Multi-arch, arch-specific compiling and dynamic function selection patch is ready.
      • It seems dual loop-unrolling gives better performance over quad-loop-unrolling.
    • iova_mode == VA not working issue is not root-caused
      • DMA mapping btw iova & pa; VPP and DPDK are using va as iova, and then do the DMA mapping.
      • However IOMMU on N1SDP requires a limited memory space, less than 40 bits?.
        • This issue will not be seen in the latest N1 Firmware. Upgrading to latest Firmware is pending.
    • Share L2/L3/ACL throughput wiz & wo L3 cache - Govind
    • Will try with L3 cache enabled to see if performance drop as flow number increasing issue is fixed or not. - Govind
      • The degradation is seen even when L3 cache is enabled.
    • Trying to make IPsec enabled with Arm platform. - Govind
    • Create Confluence page to record all the performance benchmarking data - Lijian
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang
        • Jieqiang needs 2 Intel NICs to make the test bed ready for VPP Path tests. Jieqiang and Lijian to discuss with Juraj to run the jenkins job on CentOS.

04/14/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
    • Juraj Linkes
    • Tina Tsou
    • Jieqiang Wang
  • General
  • CSIT
  • FD.io lab
  • VPP
    • Investigate bihash operations in L2 throughput are hot-spots
      • To confirm with Damjan if he has plan to rewrite l2-nodes or not - Lijian
      • To confirm CRC32 calculation is compiled with O2 or O3 options - Lijian
    • Benchmarking, profiling and comparation btw VPP and testpmd are done. Will review with team. - Lijian
    • N1SDP enablement. - Lijian
      • GCC-9.2.0 is used with "-march=armv8.2-a+crc+crypto -mtune=neoverse-n1" compiler options.
      • Multi-arch, arch-specific compiling and dynamic function selection patch is ready.
      • It seems dual loop-unrolling gives better performance over quad-loop-unrolling.
    • iova_mode == VA not working issue is not root-caused
      • DMA mapping btw iova & pa; VPP and DPDK are using va as iova, and then do the DMA mapping.
      • However IOMMU on N1SDP requires a limited memory space, less than 40 bits?.
        • This issue will not be seen in the latest N1 Firmware. Upgrading to latest Firmware is pending.
    • Share L2/L3/ACL throughput wiz & wo L3 cache - Govind
    • Will try with L3 cache enabled to see if performance drop as flow number increasing issue is fixed or not. - Govind
      • The degradation is seen even when L3 cache is enabled.
    • Trying to make IPsec enabled with Arm platform. - Govind
    • Create Confluence page to record all the performance benchmarking data - Lijian
    • Trying to reproduce the CSIT VPP device testing on local servers - Jieqiang
      • After VPP device scripts can run well on local servers, Jieqiang can investigate more features, IPv4, IPv6, Tunnel, IPSec.
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang
        • Jieqiang needs 2 Intel NICs to make the test bed ready for VPP Path tests. Jieqiang and Lijian to discuss with Juraj to run the jenkins job on CentOS.

04/07/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
    • Juraj Linkes
    • Tina Tsou
    • Jieqiang Wang
  • General
  • CSIT
  • FD.io lab
  • VPP
    • Vectorization
    • Investigate bihash operations in L2 throughput are hot-spots
      • To confirm with Damjan if he has plan to rewrite l2-nodes or not - Lijian
      • To confirm CRC32 calculation is compiled with O2 or O3 options - Lijian
    • Benchmarking, profiling and comparation btw VPP and testpmd are done. Will review with team. - Lijian
    • N1SDP enablement. - Lijian
      • GCC-9.2.0 is used with "-march=armv8.2-a+crc+crypto -mtune=neoverse-n1" compiler options.
      • Multi-arch, arch-specific compiling and dynamic function selection patch is ready.
      • It seems dual loop-unrolling gives better performance over quad-loop-unrolling.
    • iova_mode == VA not working issue is not root-caused
      • DMA mapping btw iova & pa; VPP and DPDK are using va as iova, and then do the DMA mapping.
      • However IOMMU on N1SDP requires a limited memory space, less than 40 bits?.
        • This issue will not be seen in the latest N1 Firmware. Upgrading to latest Firmware is pending.
    • Share L2/L3/ACL throughput wiz & wo L3 cache - Govind
    • Will try with L3 cache enabled to see if performance drop as flow number increasing issue is fixed or not. - Govind
      • The degradation is seen even when L3 cache is enabled.
    • Trying to make IPsec enabled with Arm platform. - Govind
    • Create Confluence page to record all the performance benchmarking data - Lijian
    • Plans
      • N1SDP performance investigation and improvement - Planned - Lijian
      • ACL plugin investigation - Planned - Govind & Lijian
      • IPsec investigation - Indicative - Govind
      • Lockless data-plane investigation by Govind in backlog
      • Continue investigating CSIT and CI management script, and run CSIT script on local servers - Jieqiang
        • Jieqiang needs 2 Intel NICs to make the test bed ready for VPP Path tests. Jieqiang and Lijian to discuss with Juraj to run the jenkins job on CentOS.

03/31/2020

03/24/2020

03/17/2020

03/10/2020

03/03/2020

02/25/2020


02/18/2020


02/11/2020


02/04/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
    • Lijian Zhang
    • Juraj Linkes
    • Tina Tsou
  • General
  • CSIT
  • FD.io lab
    • Two ThunderX2 servers are installed in Arm lab. But intel NIC can not be enumerated on one ThunderX2-01.
      • Cables for intel NICs have been ordered.
      • Universal rails will be tried with ThunderX2 servers. If it works, will send the rails to FD.io lab.
    • Script/commands to verify the NICs are ready. Will try with Mellanox NIC on ThunderX2-02 firstly - Lijian
    • Confirm the power cable requirements for the Vexxhost lab Inform about the 2 servers coming to the lab - Juraj
    • Current Configurations:
      • RAM: 256G
      • Disk: 480G SSD
      • The boxes are coming with Qlogic cards which are not supported in VPP.
    • Changes required to the servers:
      • The power cable specifications will be 4 x 6ft 14 AWG C13 to C14 Power Cables.
      • Need 2 Intel NICs XL710-QDA2 for each server.
      • If there is space for 2 more cards, we should add 2 Mellanox cards SR-IOV capability.
      • Disk size to 480G
      • Both the servers should have the NICs in the same PCIe slots. It does not matter which PCIe slots the cards are in.
      • Cables: N1, P1 to N2, P1 and so on
      • Cables for IPMI and Management port: 2
    • Will call a meeting once the ThunderX2 arrives to Dean, inviting Juraj/Vxxhost people in FD.io lab to make sure hardware ready.
    • Require and install 1G NIC for the management port because of the 1G management switch. - Lijian
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • Require one more ThunderX2 to form a normal cluster (1xThunderX + 2xThunderX2), to enable voting right for Arm servers in CSIT
    • ThunderX1
  • VPP
    • Align Arm patches with VPP release plan.
      • F0 2020-01-08 APIs frozen. Only low-risk changes accepted on main branch.
      • RC1 2020-01-15 (F0+7) Code complete, pull first release throttle branch, only bug fixes accepted on throttle train. After pull: main branch reopens for new feature / risky commits. First artifacts posted.
      • RC2 2020-01-22 (RC1+7) Second artifacts posted.
      • Formal Release 2020-01-29 (RC2+7) 20.01 release artifacts available
    • Vectorization
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Benchmarking AVF drivers on Arm servers - Jieqiang
      • VPP+DPDK (5.18Mpps/5.21Mpps) VS VPP+AVF (8.39Mpps/8.38Mpps) on ThunderX2.
      • Check if performance tests includes AVF driver or not?
    • AVF failed to create AVF interface on SMP CPU on N1SDP/Qualcomm - Jieqiang
      • Current N1SDP does not support SRIOV, so cannot run AVF on N1SDP.
      • Will try one patch to enable N1SDP board.
      • Please try AVF with Mcbin if possible.
    • Investigate bihash operations in L2 throughput are hot-spots
      • Apply prefetches and dual loop with l2_fwd node, and failed to l2_learn
      • Lock-free allocation/free give 7%-11% improvement on ThunderX2, but no improvement on x86 and CortexA72.
      • Cache misses and CRC32 calculation are possible opportunities.
        • To check cycles by applying CRC32 calculation unrolling
    • Bench-mark VPP on Dawn N1SDP board
      • Done finishing single flow with L2/L3/input-ACL on N1SDP board, will share the data.
    • Investigating bi-hash lockless implementation - Jason
      • Firstly apply make_working_copy for all bihash update operations, and then apply RCU to make look-up lock-less
    • Internal CI is not working due to Python3.6 upgrade in vpp code repository. - Jieqiang
    • EPIC for next quarter:

01/28/2020

  • Attendees
    • Govindarajan Mohandoss
    • Honnappa Nagarahalli
  • General
  • CSIT
  • FD.io lab
    • Two ThunderX2 servers are installed in Arm lab. But intel NIC can not be enumerated on one ThunderX2-01.
      • Cables for intel NICs have been ordered.
      • Universal rails will be tried with ThunderX2 servers. If it works, will send the rails to FD.io lab.
    • Script/commands to verify the NICs are ready. Will try with Mellanox NIC on ThunderX2-02 firstly - Lijian
    • Confirm the power cable requirements for the Vexxhost lab Inform about the 2 servers coming to the lab - Juraj
    • Current Configurations:
      • RAM: 256G
      • Disk: 480G SSD
      • The boxes are coming with Qlogic cards which are not supported in VPP.
    • Changes required to the servers:
      • The power cable specifications will be 4 x 6ft 14 AWG C13 to C14 Power Cables.
      • Need 2 Intel NICs XL710-QDA2 for each server.
      • If there is space for 2 more cards, we should add 2 Mellanox cards SR-IOV capability.
      • Disk size to 480G
      • Both the servers should have the NICs in the same PCIe slots. It does not matter which PCIe slots the cards are in.
      • Cables: N1, P1 to N2, P1 and so on
      • Cables for IPMI and Management port: 2
    • Will call a meeting once the ThunderX2 arrives to Dean, inviting Juraj/Vxxhost people in FD.io lab to make sure hardware ready.
    • Require and install 1G NIC for the management port because of the 1G management switch. - Lijian
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • Require one more ThunderX2 to form a normal cluster (1xThunderX + 2xThunderX2), to enable voting right for Arm servers in CSIT
    • ThunderX1
  • VPP
    • Align Arm patches with VPP release plan.
      • F0 2020-01-08 APIs frozen. Only low-risk changes accepted on main branch.
      • RC1 2020-01-15 (F0+7) Code complete, pull first release throttle branch, only bug fixes accepted on throttle train. After pull: main branch reopens for new feature / risky commits. First artifacts posted.
      • RC2 2020-01-22 (RC1+7) Second artifacts posted.
      • Formal Release 2020-01-29 (RC2+7) 20.01 release artifacts available
    • Vectorization
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Benchmarking AVF drivers on Arm servers - Jieqiang
      • VPP+DPDK (5.18Mpps/5.21Mpps) VS VPP+AVF (8.39Mpps/8.38Mpps) on ThunderX2.
      • Check if performance tests includes AVF driver or not?
    • AVF failed to create AVF interface on SMP CPU on N1SDP/Qualcomm - Jieqiang
      • Current N1SDP does not support SRIOV, so cannot run AVF on N1SDP.
      • Will try one patch to enable N1SDP board.
      • Please try AVF with Mcbin if possible.
    • Investigate bihash operations in L2 throughput are hot-spots
      • Apply prefetches and dual loop with l2_fwd node, and failed to l2_learn
      • Lock-free allocation/free give 7%-11% improvement on ThunderX2, but no improvement on x86 and CortexA72.
      • Cache misses and CRC32 calculation are possible opportunities.
        • To check cycles by applying CRC32 calculation unrolling
    • Bench-mark VPP on Dawn N1SDP board
      • Done finishing single flow with L2/L3/input-ACL on N1SDP board, will share the data.
    • Investigating bi-hash lockless implementation - Jason
      • Firstly apply make_working_copy for all bihash update operations, and then apply RCU to make look-up lock-less
    • Internal CI is not working due to Python3.6 upgrade in vpp code repository. - Jieqiang
    • EPIC for next quarter:

01/21/2020


01/14/2020

01/07/2020


12/17/2019

12/10/2019


12/03/2019

11/26/2019

11/19/2019

11/12/2019

10/29/2019

10/22/2019

10/15/2019

10/08/2019

10/01/2019

09/24/2019

09/17/2019

09/10/2019

09/03/2019

08/27/2019

08/20/2019

08/13/2019

08/06/2019

07/30/2019

  • Attendees
    • Tina Tsou
    • Honnappa Nagarahalli
    • Lijian Zhang
    • Jieqiang Wang
    • Jason Zhang
    • Juraj Linkes
    • Christian Hopps
  • General
  • CSIT
  • FD.io lab
  • VPP
    • https://tools.ietf.org/html/draft-hopps-ipsecme-iptfs-01 - From Christian
    • Align Arm patches with VPP release plan.
      • Once our work items are added to release plan, the community is forced to review the patches and provide the feedback in a timely manner.
      • Will check VPP release schedual and map with Arm Quaterly plan.
      • Note down patches in community review and align them to VPP release plan.
      • It has been challenging to do that in VPP.
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Message queue optimize it with relaxed atomic intrinsics - Lijian
    • Vectorization
        • Optimize eth_input_adv_and_flags_x4 is upstreamed and under community review.
        • The patch is also enabled for x86. Will ask maintainer to review.
    • Spinlock/read-write lock optimization - Jason
      • Refactored spinlock and added test file for spinlock. Patches are under internal review.
      • Spinlock with inner loop got improvement on both x86 and Arm.
      • Read/write lock got a little degradation with the patch.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
      • Jieqiang checked the video by Sirshak
    • Apply dual/quad optimization - Lijian
      • Benchmarking on Cortex-A72 with dpdk, ethernet-input, ip4 rewrite, tx-output nodes
      • Submitted patches on applying on dpdk-input, ethernet-input, ip4-input, ip4-rewrite nodes
      • Will do bench-marking profiling on mcbin/Bluefield.
    • Think of memory usage and optimization for smaller device/memory
    • Think about running VPP on big-endian CPU, as there's mixture usage of gcc vector extension and vector intrinsics
    • To confirm the firmware, vpp side to enable mcbin cache stashing - Honnappa
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

07/23/2019

  • Attendees
    • Tina Tsou
    • Honnappa Nagarahalli
    • Lijian Zhang
    • Jieqiang Wang
    • Jason Zhang
    • Juraj Linkes
    • Christian Hopps
  • General
  • CSIT
    • VPP Performance Test
    • Performance job is merged. https://jenkins.fd.io/view/csit/job/csit-vpp-perf-mrr-daily-master-3n-tsh
    • Trying to fix all the failures with daily test in performance test. Basically almost all the tests passed locally.
    • Only 1 out of 199 test cases failed, 8 test cases show random 'show interface' failure.
    • Some failures are related with 'show hardware'/'show interface'/'show vhost dump', time-out.
    • Adding Taishan test bed to CSIT Status: https://gerrit.fd.io/r/#/c/16850/
      • creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • VPP Path
      • Working on MAC learning test failures on Cortex-A72 server - Jieqiang
        • Enlarge duration can fix the failure, but will investigate more details.
        • Issues have been fixed in latest master branch. Investigating the details.
      • cross compilation in VPP PATH. Jira(Juraj): https://jira.fd.io/browse/CTP-3
    • VPP Device
      • VPP tap interface is not working on all Arm servers. Works on stable/1810, but not working on stable/1901.
        • Send email and current debug details to community calling for volunteer to fix it. - Lijian
        • pmalloc module test cases failed on Arm server.
      • Changes are uploaded to community gerrit.
      • vpp VMs seems to bring up well. Will work on init script and bring up vpp.
        • VM tests passed. Patches are to be submitted for community review.
        • All the patches are merged and all images are built.
      • Will confirm with Ed about where to upload VPP docker for VPP device - Juraj
      • Ed to help set up numad cluster with dual ThunderX and one ThunderX2
  • FD.io lab
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Message queue optimize it with relaxed atomic intrinsics - Lijian
    • Vectorization
        • Optimize eth_input_adv_and_flags_x4 is upstreamed and under community review.
        • The patch is also enabled for x86. Will ask maintainer to review.
    • Spinlock/read-write lock optimization - Jason
      • Refactored spinlock and added test file for spinlock. Patches are under internal review.
      • Spinlock with inner loop got improvement on both x86 and Arm.
      • Read/write lock got a little degradation with the patch.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
      • Inform MAP owner that Jieqiang will take care of MAP on VPP. - Lijian
    • Apply dual/quad optimization - Lijian
      • Benchmarking on Cortex-A72 with dpdk, ethernet-input, ip4 rewrite, tx-output nodes
      • Submitted patches on applying on dpdk-input, ethernet-input, ip4-input, ip4-rewrite nodes
      • Will do bench-marking profiling on mcbin/Bluefield.
    • Think of memory usage and optimization for smaller device/memory
    • Think about running VPP on big-endian CPU, as there's mixture usage of gcc vector extension and vector intrinsics
    • To confirm the firmware, vpp side to enable mcbin cache stashing - Honnappa
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

07/16/2019

  • Attendees
    • Tina Tsou
    • Honnappa Nagarahalli
    • Lijian Zhang
    • Jieqiang Wang
    • Jason Zhang
    • Juraj Linkes
    • Christian Hopps
  • General
  • CSIT
    • VPP Performance Test
    • Performance job is merged. https://jenkins.fd.io/view/csit/job/csit-vpp-perf-mrr-daily-master-3n-tsh
    • Adding Taishan test bed to CSIT Status: https://gerrit.fd.io/r/#/c/16850/
      • creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • VPP Path
      • Working on MAC learning test failures on Cortex-A72 server - Jieqiang
        • Enlarge duration can fix the failure, but will investigate more details.
        • Issues have been fixed in latest master branch. Investigating the details.
      • cross compilation in VPP PATH. Jira(Juraj): https://jira.fd.io/browse/CTP-3
    • VPP Device
      • VPP tap interface is not working on all Arm servers. Works on stable/1810, but not working on stable/1901.
        • Send email and current debug details to community calling for volunteer to fix it. - Lijian
      • Changes are uploaded to community gerrit.
      • vpp VMs seems to bring up well. Will work on init script and bring up vpp.
        • VM tests passed. Patches are to be submitted for community review.
        • Patch is splited into three small pieces. Two patches (kernel image for VM test/generic CSIT changes to support ThunderX2 testbed) are merged. Third patch about code changes for VM test to be merged, Arm specific code and use kernel image.
      • Will confirm with Ed about where to upload VPP docker for VPP device - Juraj
        • Docker images for both Arm and x86 are merged and available.
        • Docker image is verified on Arm server, but to verify it on x86 server also and try it in Jenkins.
      • Ed to help set up numad cluster with dual ThunderX and one ThunderX2
  • FD.io lab
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • Require one more ThunderX2 to form a normal cluster (1xThunderX + 2xThunderX2), to enable voting right for Arm servers in CSIT
      • It’s 1RU blade ThunderX2.
      • The machine will be handled by Dean’s team. Cambridge folk will set up the machine before sending it to FD.io lab.
      • The machine should have a big RAM, more than 120G and 256G preferred.
      • The machine should Three NICs (XL710-QDA2, 2x40G).
      • The script assumes the two ThunderX2 have the same NIC type, same fiber SFP type, and NICs are plugged into same PCI slots.
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Message queue optimize it with relaxed atomic intrinsics - Lijian
    • Vectorization
        • Optimize eth_input_adv_and_flags_x4 is upstreamed and under community review.
        • The patch is also enabled for x86. Will ask maintainer to review.
    • Spinlock/read-write lock optimization - Jason
      • Refactored spinlock and added test file for spinlock. Patches are under internal review.
      • Spinlock with inner loop got improvement on both x86 and Arm.
      • Read/write lock got a little degradation with the patch.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Apply dual/quad optimization - Lijian
      • Benchmarking on Cortex-A72 with dpdk, ethernet-input, ip4 rewrite, tx-output nodes
      • Submitted patches on applying on dpdk-input, ethernet-input, ip4-input, ip4-rewrite nodes
      • Will do bench-marking profiling on mcbin.
    • Think of memory usage and optimization for smaller device/memory
    • Think about running VPP on big-endian CPU, as there's mixture usage of gcc vector extension and vector intrinsics
    • To confirm the firmware, vpp side to enable mcbin cache stashing - Honnappa
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

07/09/2019

  • Attendees
    • Tina Tsou
    • Honnappa Nagarahalli
    • Lijian Zhang
    • Jieqiang Wang
    • Jason Zhang
    • Juraj Linkes
    • Christian Hopps
  • General
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server: https://jira.fd.io/browse/VPP-1569
    • Adding Taishan test bed to CSIT Status: https://gerrit.fd.io/r/#/c/16850/
      • creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • VPP Path
    • VPP Device
      • VPP tap interface is not working on all Arm servers. Works on stable/1810, but not working on stable/1901.
        • Send email and current debug details to community calling for volunteer to fix it. - Lijian
      • Changes are uploaded to community gerrit.
      • vpp VMs seems to bring up well. Will work on init script and bring up vpp.
        • VM tests passed. Patches are to be submitted for community review.
      • Will confirm with Ed about where to upload VPP docker for VPP device - Juraj
        • Docker images for both Arm and x86 are merged and available.
      • Ed to help set up numad cluster with dual ThunderX and one ThunderX2
  • FD.io lab
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • Discuss requiring another ThunderX2 1U blade with Pravin and Tina - Lijian
      • Update the current status to Pravin. - Lijian
      • The machine will be handled by Jingjing's team. Cambridge folk will set up the machine before sending it to FD.io lab.
      • Require a bigger than 120G RAM, prefer 256G
      • Three NICs and each has two ports.
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Message queue optimize it with relaxed atomic intrinsics - Lijian
    • Vectorization
        • Optimize eth_input_adv_and_flags_x4 is upstreamed and under community review.
    • Spinlock/read-write lock optimization - Jason
      • Refactored spinlock and added test file for spinlock. Patches are under internal review.
      • Spinlock with inner loop got improvement on both x86 and Arm.
      • Read/write lock got a little degradation with the patch.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Apply dual/quad optimization - Lijian
      • Benchmarking on Cortex-A72 with dpdk, ethernet-input, ip4 rewrite, tx-output nodes
      • Submitted patches on applying on dpdk-input, ethernet-input, ip4-input, ip4-rewrite nodes
      • Will do bench-marking profiling on mcbin.
    • Think of memory usage and optimization for smaller device/memory
    • Think about running VPP on big-endian CPU, as there's mixture usage of gcc vector extension and vector intrinsics
    • To confirm the firmware, vpp side to enable mcbin cache stashing - Honnappa
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective


07/02/2019

  • Attendees
    • Tina Tsou
    • Honnappa Nagarahalli
    • Lijian Zhang
    • Jieqiang Wang
    • Jason Zhang
    • Juraj Linkes
  • General
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server: https://jira.fd.io/browse/VPP-1569
    • Adding Taishan test bed to CSIT Status: https://gerrit.fd.io/r/#/c/16850/
      • creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • VPP Path
    • VPP Device
      • VPP tap interface is not working on all Arm servers. Works on stable/1810, but not working on stable/1901.
        • Send email and current debug details to community calling for volunteer to fix it. - Lijian
      • vpp VMs seems to bring up well. Will work on init script and bring up vpp.
      • Will confirm with Ed about where to upload VPP docker for VPP device - Juraj
      • Set up numad cluster with dual ThunderX and one ThunderX2
  • FD.io lab
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • Discuss requiring another ThunderX2 1U blade with Pravin and Tina - Lijian
      • Update the current status to Pravin. - Lijian
      • The machine will be handled by Jingjing's team. Cambridge folk will set up the machine before sending it to FD.io lab.
      • Require a bigger than 120G RAM, prefer 256G
      • Three NICs and each has two ports.
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Message queue, remove atomic intrinsics and use lock version only - Lijian
      • Have verified removing atomic intrinsics in message_queue alloc/free APIs, and require confirmation from Florin.
    • Vectorization
        • Optimize eth_input_adv_and_flags_x4 is upstreamed and under community review.
    • Spinlock/read-write lock optimization - Jason
      • Refactored spinlock and added test file for spinlock. Patches are under internal review.
    • Fix ip4_forward compiling - Jason
      • Will check gerrit CI/CD related with that patch. Check why it's not warning in gerrit Jenkins.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Spread dual/quad optimization - Lijian
      • Benchmarking on Cortex-A72 with dpdk, ethernet-input, ip4 rewrite, tx-output nodes
      • Will do bench-marking profiling on mcbin.
    • Think of memory usage and optimization for smaller device/memory
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

06/25/2019

  • Attendees
    • Tina Tsou
    • Honnappa Nagarahalli
    • Lijian Zhang
    • Jieqiang Wang
    • Jason Zhang
    • Juraj Linkes
  • General
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server: https://jira.fd.io/browse/VPP-1569
    • Adding Taishan test bed to CSIT Status: https://gerrit.fd.io/r/#/c/16850/
      • creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • VPP Path
    • VPP Device
      • VPP tap interface is not working on all Arm servers. Works on stable/1810, but not working on stable/1901.
      • Crypto test cases, will use dpdk driver if configured, native-vpp implementation, fall back to openSSL
        • Will try Crypto test cases next week - Juraj
      • Juraj to send Lijian the details of vpp VMs, Lijian will confirm internally
  • FD.io lab
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • Discuss requiring another ThunderX2 1U blade with Pravin and Tina - Lijian
      • Firstly will sponsor the machine
      • The machine will be handled by Jingjing's team. Cambridge folk will set up the machine before sending it to FD.io lab.
      • Require a bigger than 120G RAM, prefer 256G
      • Three NICs and each has two ports.
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Message queue, remove atomic intrinsics and use lock version only - Lijian
      • Have verified removing atomic intrinsics in message_queue alloc/free APIs, and require confirmation from Florin.
    • Vectorization
        • Optimize eth_input_adv_and_flags_x4 is upstreamed and under community review.
    • Spinlock optimization - Jason
      • Refactored spinlock and added test file for spinlock. Patches are under internal review.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Spread dual/quad optimization - Lijian
      • Benchmarking on Cortex-A72 with dpdk, ethernet-input, ip4 rewrite, tx-output nodes
      • Will do bench-marking profiling on mcbin.
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

06/18/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina Tsou
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj Linkes
  • General
  • CSIT
  • FD.io lab
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
      • Will confirm with Florin to remove atomic intrinsics in message_queue alloc/free APIs
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - Upstreamed.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Investigate hyperscan plugin in VPP - Sirshak
    • Spread qual/quad optimization - ethernet-input
    • Redo perf/MAP profiling/bench-marking
      • DPI plugin?
    • EPIC for next quarter:
      • Apply dual/quad optimization on more data path nodes
      • Investigate and optimize VPP hash and bihash library
      • VPP translation overhead analysis btw Mbuf and VLIB buffer ENTNET-1293
      • VPP Memif performance analysis and optimization ENTNET-1292
      • VPP l3fwd performance analysis and optimization ENTNET-751
      • Using MAP with VPP ENTNET-1288

06/11/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina Tsou
    • Lijian Zhang
    • Jieqiang Wang
    • Juraj
  • General
  • CSIT
  • FD.io lab
    • Require two ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
      • Will confirm with Florin to remove atomic intrinsics in message_queue alloc/free APIs
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - Upstreamed.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Investigate hyperscan plugin in VPP - Sirshak
    • Spread qual/quad optimization - ethernet-input
    • Redo perf/MAP profiling/bench-marking
      • DPI plugin?
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

06/04/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina Tsou
    • Lijian Zhang
    • Jieqiang Wang
    • Stan
  • General
  • CSIT
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - Upstreamed.
    • MAP with VPP - error is resolved. Sort of working. Record the details.
    • Investigate hyperscan plugin in VPP - Sirshak
      • DPI plugin?
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

05/28/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina
    • Lijian Zhang
    • Vijay (vijayakumar.rajamanickam@nokia.com)
  • General
  • CSIT
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - internal Review.
    • MAP with VPP - Tried internal Patch still failing. Continuing to work on it.
    • Investigate hyperscan plugin in VPP - Sirshak
      • DPI plugin?
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

05/21/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina
    • Lijian Zhang
    • Vijay (vijayakumar.rajamanickam@nokia.com)
  • General
  • CSIT
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - internal Review.
    • MAP with VPP - Tried internal Patch still failing. Continuing to work on it.
    • Investigate hyperscan plugin in VPP - Sirshak
      • DPI plugin?
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

05/14/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina
    • Lijian Zhang
    • Vijay (vijayakumar.rajamanickam@nokia.com)
  • General
  • CSIT
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
    • VPP generic distro package building patch - Patch updated. Require Damjan's follow up review.
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - internal Review.
    • MAP with VPP - Tried internal Patch still failing. Continuing to work on it.
    • Investigate hyperscan plugin in VPP - Sirshak
      • DPI plugin?
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

05/07/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina
    • Lijian Zhang
    • Vijay (vijayakumar.rajamanickam@nokia.com)
  • General
  • CSIT
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
      • Investigate ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0) in dpdk_device_input.
    • VPP generic distro package building patch - Patch updated Damjan's follow up review required.
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • Vectorization in esp-encrypt, optimize memcpy_le. Upstreamed(https://gerrit.fd.io/r/#/c/18398/). - Lijian
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - internal Review.
    • MAP(Arm Proprietary Performance Analysis Tool) with VPP - Tried internal Patch still failing. Continuing to work on it.
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

04/30/2019

  • Attendees
    • Sirshak Das
    • Honnappa Nagarahalli
    • Tina
  • General
  • CSIT
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
  • VPP
    • VPP host-stack Hotspots
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
      • Investigate ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0) in dpdk_device_input.
    • VPP generic distro package building patch - Patch updated Damjan's follow up review required.
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - ongoing - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • VPP machiatto bin showing some unstable performance.
    • Vectorization
      • Vectorization in esp-encrypt, optimize memcpy_le. Upstreamed(https://gerrit.fd.io/r/#/c/18398/). - Lijian
      • ethernet-input causes performance drop on AArch64.
        • There's performance drop issue after the ethernet-input optimization. The major reason is after the refactor, if promiscuous is enabled on NIC, all traffic from the NIC will fall into so-called slow path.
        • A vectorized patch to optimize eth_input_adv_and_flags_x4 is under internal review.
    • TAS patch - internal Review.
    • MAP(Arm Proprietary Performance Analysis Tool) with VPP - Tried internal Patch still failing. Continuing to work on it.
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective

04/23/2019

  • Attendees
    • Sirshak Das
    • Lijian Zhang
    • Juraj Linkeš
    • Vijay
    • Nitin
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • List all the blockers on aarch64 in CSIT wiki page - Done - Stan or Juraj
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • QSFP+ is available and working now.
        • Juraj to resend email to Mahamad about the details, including Sirshak and Tina
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
        • Confirm with Gorka if their mcbin can support docker. If yes, then ask them to provide image with their latest kernel/file system/dtd
      • thunderx2: Status: server is set up. Management connection works. Intel NICs are well connected. Will prepare the server for VPP device testing. Now is working on containers for VPP device. Will probably be able to run VPP device tests manually this week.
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Management connection thru QSFP+ switch is working now. Moving VPP device work to ThunderX1 blade servers.
      • Will use these four new ThunderX1 servers for CI, Genkins to replace the previous three old ThunderX1 servers.
      • These four ThunderX1 blade are not identical. The first one has two numa nodes, and other three blades have one numa node.
      • Investigate why these three blades have only one numa node - Juraj
  • VPP
    • Investigate session_queue_node_fn/vlib_worker_loop.
      • Decrease or remove ATOMIC_ACQUIRE atomics in foreach_device_and_queue
      • Investigate ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0) in dpdk_device_input
    • Rebase VPP distro package building patch; contact Damjan in slack; Talk with Damjan in vpp meeting - Lijian & Sirshak
    • Investigating message queue, understand use case with svm queue, talk the ideas with Florin - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Perf degradation is fixed. Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • Octeon-Tx Status(Sirshak): Done by Malvika. Running DPDK sample apps on it now.
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
    • Vectorization
    • TAS patch will be ready soon (Sirshak)
    • MAP with VPP is ongoing - Sirshak
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective
  • Action Items - Last Week
  • Action Items - Next Week

04/16/2019

  • Attendees
    • Sirshak Das
    • Lijian Zhang
    • Juraj Linkeš
    • Vijay
    • Nitin
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • List all the blockers on aarch64 in CSIT wiki page - Done - Stan or Juraj
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • QSFP+ is available and working now.
        • Juraj to resend email to Mahamad about the details, including Sirshak and Tina
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
        • Confirm with Gorka if their mcbin can support docker. If yes, then ask them to provide image with their latest kernel/file system/dtd
      • thunderx2: Status: server is set up. Management connection works. Intel NICs are well connected. Will prepare the server for VPP device testing. Now is working on containers for VPP device. Will probably be able to run VPP device tests manually this week.
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Management connection thru QSFP+ switch is working now. Moving VPP device work to ThunderX1 blade servers.
      • Will use these four new ThunderX1 servers for CI, Genkins to replace the previous three old ThunderX1 servers.
      • These four ThunderX1 blade are not identical. The first one has two numa nodes, and other three blades have one numa node.
      • Investigate why these three blades have only one numa node - Juraj
  • VPP
    • Investigate session_queue_node_fn/vlib_worker_loop. - https://jira.arm.com/browse/ENTNET-1179 - Done
      • Will create two Jira tickets to track the findings. - Lijian
    • Rebase VPP distro package building patch; contact Damjan in slack; Talk with Damjan in vpp meeting - Lijian & Sirshak
    • Investigating message queue - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Perf degradation is fixed. Investigating performance degradation on Bluefield - Sirshak
        • Will resume Taishan host-stack setup - Lijian
      • Octeon-Tx Status(Sirshak): Done by Malvika. Running DPDK sample apps on it now.
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
    • Vectorization
    • EPIC for next quarter:
      • ethernet-input - Planned (Lijian). will implement for aarch64 128bits only
      • Message Queue - Planned (Lijian)
      • VPP svm_fifo patch performance optimization on A72 cores – Planned (Sirshak)
      • TAS patch (Sirshak)
      • MAP with VPP - Planned (Sirshak)
      • Roadmap for TCP optimization
        • Timer implementation - (Sirshak) - Indicative
        • perf analysis - Planned (Sirshak)
          • TCP state machine from weak memory model perspective
  • Action Items - Last Week
  • Action Items - Next Week

04/09/2019

  • Attendees
    • Sirshak Das
    • Lijian Zhang
    • Juraj Linkeš
    • Nitin
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Honnappa Nagarahalli
  • General
  • CSIT
    • VPP Performance Test
    • List all the blockers on aarch64 in CSIT wiki page - Done - Stan or Juraj
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • QSFP+ is available and working now.
        • Juraj to resend email to Mahamad about the details, including Sirshak and Tina
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
        • Confirm with Gorka if their mcbin can support docker. If yes, then ask them to provide image with their latest kernel/file system/dtd
      • thunderx2: Status: server is set up. Management connection works. Intel NICs are well connected. Will prepare the server for VPP device testing. Now is working on containers for VPP device. Will probably be able to run VPP device tests manually this week.
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Management connection thru QSFP+ switch is working now. Moving VPP device work to ThunderX1 blade servers.
      • Will use these four new ThunderX1 servers for CI, Genkins to replace the previous three old ThunderX1 servers.
      • These four ThunderX1 blade are not identical. The first one has two numa nodes, and other three blades have one numa node.
      • Investigate why these three blades have only one numa node - Juraj
  • VPP
  • VPP Hoststack
    • Investigate session_queue_node_fn/vlib_worker_loop. - https://jira.arm.com/browse/ENTNET-1179 - Done
    • Rebase VPP distro package building patch; contact Damjan in slack; Talk with Damjan in vpp meeting - Lijian & Sirshak
    • Investigating message queue - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Perf degradation is fixed. Investigating performance degradation on Bluefield - Sirshak
      • Octeon-Tx Status(Sirshak): Done by Malvika. Running DPDK sample apps on it now.
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
    • Vectorization
      • Vecterization in esp-encrypt, https://gerrit.fd.io/r/#/c/18398/ - Get improvement on ThunderX/OcteonTX/Taishan, but degradation on ThunderX2 - Lijian
      • ethernet-input - will implement for aarch64 128bits only
      • Create vectorization specific EPIC - Lijian
  • Action Items - Last Week
  • Action Items - Next Week

04/02/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Nitin
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • After assigned dedicated CPU processors for VPP main/VPP worker/iperf3 server, both ThunderX2 and Taishan Server VPP hoststack give better performance compared with Linux stack.
    • Investigate session_queue_node_fn/message queue data structure. - Investigating the source code
    • Review https://gerrit.fd.io/r/#/c/18398/ - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch. Upstreamed. Perf degradation is fixed.
      • Octeon-Tx Status(Sirshak): Done by Malvika. Running DPDK on it now.
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • List all the blockers on aarch64 in CSIT wiki page - Stan or Juraj
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Both binaries and packages built with generic option by default, and provide Makefile variable NATIVE_OPTIMIZE=Y for end user to build native optimized images.
      • Prepare email and a draft patch asking comments from community - Lijian
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • QSFP+ is available and working now.
        • Juraj to resend email to Mahamad about the details, including Sirshak and Tina
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
        • Confirm with Gorka if their mcbin can support docker. If yes, then ask them to provide image with their latest kernel/file system/dtd
      • Confirm if Jianlin's board has the exactly same plugable switches with Juraj's boards - Lijian
      • thunderx2: Status: server is set up. Management connection works. Intel NICs are well connected. Will prepare the server for VPP device testing. Now is working on containers for VPP device. Will probably be able to run VPP device tests manually this week.
  • FD.io lab
    • Require one ThunderX2(currently only one thunderX2 in the lab) in FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Management connection thru QSFP+ switch is working now. Moving VPP device work to ThunderX1 blade servers.
      • Will use these four new ThunderX1 servers for CI, Genkins to replace the previous three old ThunderX1 servers.
      • These four ThunderX1 blade are not identical. The first one has two numa nodes, and other three blades have one numa node.
      • Investigate why these three blades have only one numa node - Juraj
  • VPP
    • Write description/expectation about the two NEON related patch - Lijian
    • Investigating performance degradation on CortexA72 - Sirshak
    • Message queue - Sirshak
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560 - To close it.
    • Vectorization
      • ethernet-input - no progress yet
    • 128B cache line size
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week

03/26/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Nitin
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • After assigned dedicated CPU processors for VPP main/VPP worker/iperf3 server, both ThunderX2 and Taishan Server VPP hoststack give better performance compared with Linux stack.
    • Investigate session_queue_node_fn/message queue data structure. - Investigating the source code
    • Review https://gerrit.fd.io/r/#/c/18398/ - Lijian
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch. Upstreamed. Perf degradation is fixed.
      • Octeon-Tx Status(Sirshak): Done by Malvika. Running DPDK on it now.
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Both binaries and packages built with generic option by default, and provide Makefile variable NATIVE_OPTIMIZE=Y for end user to build native optimized images.
      • Prepare email and a draft patch asking comments from community - Lijian
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • QSFP+ is available and working now.
        • Juraj to resend email to Mahamad about the details, including Sirshak and Tina
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
        • Confirm with Gorka if their mcbin can support docker. If yes, then ask them to provide image with their latest kernel/file system/dtd
      • Confirm if Jianlin's board has the exactly same plugable switches with Juraj's boards - Lijian
      • thunderx2: Status: server is set up. Management connection works. Intel NICs are well connected. Will prepare the server for VPP device testing. Now is working on containers for VPP device. Will probably be able to run VPP device tests manually this week.
  • FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Management connection thru QSFP+ switch is working now. Moving VPP device work to ThunderX1 blade servers.
      • Will use these four new ThunderX1 servers for CI, Genkins to replace the previous three old ThunderX1 servers.
      • These four ThunderX1 blade are not identical. The first one has two numa nodes, and other three blades have one numa node.
      • Investigate why these three blades have only one numa node - Juraj
  • VPP
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560 - To close it.
    • Vectorization
      • ethernet-input - no progress yet
    • 128B cache line size
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week

03/19/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • After assigned dedicated CPU processors for VPP main/VPP worker/iperf3 server, both ThunderX2 and Taishan Server VPP hoststack give better performance compared with Linux stack.
    • vlib_worker_loop and session_queue_node_fn are two major hot-spots. - Just started
    • Enable NEON instruction in Buffer pool free function. Patch is committed.
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch. Upstreamed, but still working on issues, e.g., performance degradation
      • Octeon-Tx Status(Sirshak): yet to try steps from gorka for usb ubuntu rootfs installation. - Done by Malvika.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
      • Prepare email and a draft patch asking comments from community - Lijian
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
        • Juraj to resend email to Mahamad about the details, including Sirshak and Tina
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
        • Confirm with Gorka if their mcbin can support docker. If yes, then ask them to provide image with their latest kernel/file system/dtd
      • Confirm if Jianlin's board has the exactly same plugable switches with Juraj's boards - Lijian
      • thunderx2: Status: Talk to edk about deployment strategy with 1-node. Also blocked by QSFP+ issue.
  • FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Juraj setup call with LF people. Status: Done.
    • ThunderX2
      • Cables: Sent. Juraj to open another tkt for wiring the ThunderX2.
  • VPP
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560 - In internal review.
    • Vectorization
    • 128B cache line size
      • VPP image with 128B cache line size crashed on ThunderX2 - Cannot reproduce crash with my setup
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week
    • Commit VPP distro making patch - Lijian
    • Plugin 25G NIC Taishan server, and connect the 25G ports to x86 25G NIC - Lijian
    • Follow Jianlin's suggestion, update Uboot and Kernel, and then sync up with Juraj - Lijian

03/12/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
    • Tina to update the meeting notice.
  • VPP Hoststack
    • After assigned dedicated CPU processors for VPP main/VPP worker/iperf3 server, both ThunderX2 and Taishan Server VPP hoststack give better performance compared with Linux stack.
    • vlib_worker_loop and session_queue_node_fn are two major hot-spots. - No progress
    • Enable NEON instruction in Buffer pool free function. Patch is committed.
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch. All test failures are resolved.
      • Octeon-Tx Status(Sirshak): yet to try steps from gorka for usb ubuntu rootfs installation. - Switched to Malvika.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
      • Prepare email and a draft patch asking comments from community - Lijian
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
      • Confirm if Jianlin's board has the exactly same plugable switches with Juraj's boards - Lijian
      • thunderx2: Status: Talk to edk about deployment strategy with 1-node.
  • FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Juraj setup call with LF people. Status: Done.
    • ThunderX2
      • Cables: Sent. Juraj to open another tkt for wiring the ThunderX2.
  • VPP
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560 - In internal review.
    • Vectorization
    • 128B cache line size
      • VPP image with 128B cache line size crashed on ThunderX2
    • thunderx2 crashing - No update
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week
    • Commit VPP distro making patch - Lijian
    • Plugin 25G NIC Taishan server, and connect the 25G ports to x86 25G NIC - Lijian
    • Follow Jianlin's suggestion, update Uboot and Kernel, and then sync up with Juraj - Lijian

03/05/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • After assigned dedicated CPU processors for VPP main/VPP worker/iperf3 server, both ThunderX2 and Taishan Server VPP hoststack give better performance compared with Linux stack.
    • vlib_worker_loop and session_queue_node_fn are two major hot-spots. - No progress
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch. All test failures are resolved.
      • Octeon-Tx Status(Sirshak): yet to try steps from gorka for usb ubuntu rootfs installation. - Switched to Malvika.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
      • Confirm if Jianlin's board has the exactly same plugable switches with Juraj's boards - Lijian
      • thunderx2: Status: Talk to edk about deployment strategy with 1-node.
  • FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Juraj setup call with LF people. Status: Done.
    • ThunderX2
      • Cables: Sent. Juraj to open another tkt for wiring the ThunderX2.
  • VPP
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560 - No progress
      • Investigate with latest VPP code on x86 server - Lijian - Send emails to vpp-dev mailor if there's problem. Will not put much effort.
    • Vectorization
      • ethernet-input
      • buffer pools
    • 128B cache line size
      • Will try this on Taishan server - Slightly performance degradation with 128 bytes cache line
    • thunderx2 crashing - No update
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week

02/26/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • After assigned dedicated CPU processors for VPP main/VPP worker/iperf3 server, both ThunderX2 and Taishan Server VPP hoststack give better performance compared with Linux stack.
    • el0_sys hot-spot on Taishan D05 only, no plan to fix it.
    • vlib_worker_loop and session_queue_node_fn are two major hot-spots. - No progress
    • memcpy optimization
      • memcpy patch verification on taishan by khem l3 forwarding usecase- Lijian Status(khem): No updates.
      • memcpy patch consumes more clocks in OcteonTX2 - updated by Nitin.
      • Stopped working on this patch.
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch. Test failure on SCTP, not root-caused yet.
      • Octeon-Tx Status(Sirshak): yet to try steps from gorka for usb ubuntu rootfs installation. - Switched to Marvikar
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config, almost done with patch https://gerrit.fd.io/r/#/c/16837/ - Done
      • b. merging CSIT patch. - Closing done
      • c. creating a job. - Everything is ready except the docker image
    • Target: master trending job - firstly create trending graph from daily data; then create release report(require some manual work)
    • Add license header/copy right to scripts - Sirshak/Honnappa to confirm with Andy Waffa
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
      • Confirm if Jianlin's board has the exactly same plugable switches with Juraj's boards - Lijian
      • thunderx2: Status: Talk to edk about deployment strategy with 1-node.
  • FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Juraj setup call with LF people. Status: Done.
    • ThunderX2
      • Cables: Sent. Juraj to open another tkt for wiring the ThunderX2.
  • VPP
    • Buffer Pools per NUMA
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560
      • Investigate with latest VPP code on x86 server - Lijian - Send emails to vpp-dev mailor if there's problem. Will not put much effort.
    • Vectorization
      • ethernet-input
      • buffer pools
    • 128B cache line size
      • Will try this on Taishan server - Slightly performance degradation with 128 bytes cache line
    • Qualcomm no change iperf3
    • thunderx2 crashing - No update
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week

02/19/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • After assigned dedicated CPU processors for VPP main/VPP worker/iperf3 server, both ThunderX2 and Taishan Server VPP hoststack give better performance compared with Linux stack.
    • memcpy optimization
      • memcpy patch verification on taishan by khem l3 forwarding usecase- Lijian Status(khem): No updates.
      • memcpy patch consumes more clocks in OcteonTX2 - updated by Nitin.
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch.
      • Octeon-Tx Status(Sirshak): yet to try steps from gorka for usb ubuntu rootfs installation.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • VPP running on Arm side, x86 iperf3 client observes unstable performance rate.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible. https://jira.fd.io/browse/VPP-1566
    • Installed VPP crashed on Taishan server, https://jira.fd.io/browse/VPP-1569
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
    • Target: master trending job
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
      • Confirm if Jianlin's board has the exactly same plugable switches with Juraj's boards - Lijian
      • thunderx2: Status: Talk to edk about deployment strategy with 1-node.
  • FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Juraj setup call with LF people. Status: Done.
    • ThunderX2
      • Cables: Sent. Juraj to open another tkt for wiring the ThunderX2.
  • VPP
    • Buffer Pools per NUMA
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560
    • 1GB page taking long time Status: fixed.
      • Investigate with latest VPP code on x86 server
    • Vectorization
      • ethernet-input
      • buffer pools
      • memcpy
    • 128B cache line size
      • Will try this on Taishan server - Lijian
    • Qualcomm no change iperf3
    • thunderx2 crashing - No update
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week

02/11/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • memcpy optimization
      • memcpy patch verification on taishan by khem l3 forwarding usecase- Lijian Status(khem): No updates.
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation Status(Sirshak): no updates.
      • svm_fifo: Status(Sirshak): Working on fixing VPP Path errors from svm_fifo patch.
      • Octeon-Tx Status(Sirshak): yet to try steps from gorka for usb ubuntu rootfs installation.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
  • CSIT
    • VPP Performance Test
    • Package Installation error Status(Juraj): interfaces configured numa node 2,3 are not visible.
    • Estimates from Khem and Stan/Juraj. Status: https://gerrit.fd.io/r/#/c/16850/
      • a. Host Config
      • b. merging CSIT patch.
      • c. creating a job.
    • Target: master trending job
    • VPP Path
    • VPP Device
      • thunderx Status: 1-node topology was rewired because of QSFP+ switch.
      • mcbin: Kernel Migration on mcbin. Status: was able to update uboot but not boot the new kernel. Jialin suggested different boot parameters, juraj yet to try.
      • thunderx2: Status: Talk to edk about deployment strategy with 1-node.
  • FD.io lab
    • ThunderX1
      • QSFP+ switch for ThunderX1 Status: ONL OS to be installed on QSFP+ switch.
      • Juraj setup call with LF people. Status: Done.
    • ThunderX2
      • Cables: Sent. Juraj to open another tkt for wiring the ThunderX2.
  • VPP
    • Buffer Pools per NUMA
    • Verify effects and make NEON changes Jira: https://jira.fd.io/browse/VPP-1560
    • 1GB page taking long time Status: fixed.
    • Vectorization
      • ethernet-input
      • buffer pools
      • memcpy
    • 128B cache line size
    • Qualcomm no change iperf3
    • thunderx2 crashing
    • Taishan/A72 Status: Khem to try 128B cache line on taishan (performance difference).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week

02/05/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Gorka
    • Fede
    • Honnappa Nagarahalli
  • General
  • VPP Hoststack
    • memcpy optimization
      • Check optimized memory copy version are deployed on Taishan and ThunderX2 during runtime - Lijian
      • Send memcpy patch to Khem and Fede for further verification - Lijian Status: fede: small improvement in mcbin with iperf3, khem to try them with l3 forwarding
    • iperf3 performance with Hoststack.
      • ip4_local_inline quad loop under investigation
      • Working on svm_fifo alternate version with front and back pointers synchronized instead of cursize.
    • Verifying per NUMA node buffer pool https://gerrit.fd.io/r/#/c/16638/
      • sirshak create jira id in fd.io jira. https://jira.fd.io/browse/VPP-1560
      • Hanging of VPP is actually VPP taking a lot of time to allocate 400K chunks for 1GB - Damjan has this in his todo list
      • gcc-8 compilation still fails on ARM.
      • Octeon-Tx failure. Status: unknown
    • Gorka is trying some optimal configs for VCL. Status: no updates.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • OcteonTx boots to buildroot with no dhclient hence an impasse. Still not clear how to use USB stick.
  • CSIT
    • VPP Path
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx. Status: no updates.
      • Kernel Migration on mcbin. Status:
      • ThunderX2:
    • VPP Performance Test
      • Stan and Khem to come up with a summary of current status and an estimate of at least upstreaming basic L2/L3 performance suites.
      • Juraj to come up with a solution for NUMA node anamoly in Taishan.
      • https://gerrit.fd.io/r/#/c/16850/ Status: Juraj has a version all ready to work. Package installation blocker.
      • Package installation error Status: Juraj to investigate logs.
  • FD.io lab
    • ThunderX1 -
      • New QSFP+ switch for ThunderX1 is available now: QSFP+ to be connected SFP+ switch.
      • Juraj to setup a call with LF folks on.
    • ThunderX2 -
      • Andy still waiting cables.
      • Juraj to remind Andy of when the cable will be available.
      • Juraj to follow up on ssh connectivity to thunderx2.
  • VPP
    • Vectorization
      • [Lijian] Macro benchmarking on ThunderX2/Centriq(4%)/Taishan D05(10%) is done, data is updated into Jira. Code is in internal review.
      • [Lijian] Check if setting default cache line size to 128 will degradate thru-put on Taishan/Qualcomm/ThunderX2
        • no perf diff in Qualcomm
        • vpp crashes on thunderx2
        • waiting for results on A72 (Taishan)
      • [Sirshak] on ethernet-input node, investigate vectorized buffer index, Damjan's per numa node buffer pool patch. Status: No updates
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
  • Action Items - Next Week

01/29/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Lijian Zhang
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Garcia
    • Manuel
    • Gorka
    • Fede
    • Honnappa Nagarahalli
    • John Ddigilio
  • General
  • VPP Hoststack
    • TaiShan Server with Debian distro crashed with command of 'ip probe-neighbor' when doing VPP hoststack with iperf3
    • With 64 bytes packets, on ThunderX2, 10G NIC, VPP hoststack bandwidth is about 1/2 of Linux Kernel stack.
    • With 64 bytes packets, on Taishan, 10G NIC, VPP hoststack bandwidth is about 2x of Linux Kernel stack.
    • Memory copy patch gives 4% improvement on VPP hoststack on Taishan server.
    • Check optimized memory copy version are deployed on Taishan and ThunderX2 during runtime - Lijian
    • Send memcopy patch to Khem and Fede for further verification - Lijian
    • iperf3 performance with Hoststack.
      • Investigating lock-less fifo, ip4_local_forward node and memory reordering for VPP hoststack - Sirshak
    • Sirshak is trying to set up one fd.io setup for everyone debugging VPP hoststack.
    • Verifying https://gerrit.fd.io/r/#/c/16638/ - Suppose to give better performance, but VPP hang with this patch on some Arm machines.
    • Gorka is trying some optimal configs for VCL. - No Updates.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • OcteonTX is received in ARM lab. Will boot it up firstly and then start doing profiling with it.
  • FD.io lab
    • ThunderX1 -
      • New Arista switch for ThunderX1 is available now. Gathering details that required by LF lab before sending the switch to CSIT lab. - Juraj - Andy will try to send the switch to CSIT this Thursday.
    • ThunderX2 -
      • Cable type is confirmed. Procurement is in the process.
      • Juraj to remind Andy of when the cable will be available.
      • Require access to these servers in FD.io lab. Anton gives the IP to access them.(ADMIN/ADMIN)
  • CSIT
    • VPP Path
      • So far so good.
      • ARM CI results are overwritten by x86 machines. Should be a Jenkin issue. Monitor if this corner will happen again. - Juraj
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx. Basic skeleton of docker topology done. Moving on to deploying the scripts. Juraj is able to modify, execute the scripts in Container. Things to fix is scripts for 1-link 1-node topology and interfaces binding to VPP. Is able to run successfully a traffic test.
      • Kernel Migration on mcbin. Juraj is able to build all the images, but got kernel panic. Try with the more recent uBoot version. Tried latest uBoot image, but still has the same issue.
      • Juraj to investigate further work once ThunderX2 is available.
    • VPP Performance Test
  • VPP
    • Vectorization
      • [Lijian] Macro benchmarking on ThunderX2/Centriq(4%)/Taishan D05(10%) is done, data is updated into Jira. Code is in internal review.
      • [Lijian] Check if setting default cache line size to 128 will degradate thru-put on Taishan/Qualcomm/ThunderX2
      • [Sirshak] on ethernet-input node, investigate vectorized buffer index.
    • Memory Ordering
      • [Sirshak] svm_fifo lockless alternate algorithm for SPSC. - In internal review.
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. - Bootstrap script common.sh working and checking the test cases.
    • [Juraj] IP4 reassembly and GBP failures are fixed. Patches to enable them two are merged. No any test cases in blacklist for Aarch64 machine.
    • [Juraj] Kernel Migration on mcbin. Juraj is able to build all the images.
  • Action Items - Next Week
    • [Sirshak] -

01/22/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Lijian Zhang
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Garcia
    • Manuel
    • Gorka
    • Fede
    • Honnappa Nagarahalli
    • John Ddigilio
  • General
  • VPP Hoststack
    • TaiShan Server with Debian distro crashed with command of 'ip probe-neighbor' when doing VPP hoststack with iperf3
    • With 64 bytes packets, on ThunderX2, 10G NIC, VPP hoststack bandwidth is about 1/4 of Linux Kernel stack.
    • iperf3 performance with Hoststack.
      • Investigating lock-less fifo, ip4_local_forward node and memory reordering for VPP hoststack - Sirshak
    • Sirshak is trying to set up one fd.io setup for everyone debugging VPP hoststack.
    • Gorka is trying some optimal configs for VCL. - No Updates.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • OcteonTX is received in ARM lab. Will boot it up firstly and then start doing profiling with it.
  • FD.io lab
    • ThunderX1 -
      • New Arista switch for ThunderX1 is available now. Gathering details that required by LF lab before sending the switch to CSIT lab. - Juraj - Andy will try to send the switch to CSIT this Thursday.
    • ThunderX2 -
      • Cable type is confirmed. Procurement is in the process.
      • Require access to these servers in FD.io lab.
  • CSIT
    • VPP Path
      • So far so good.
      • ARM CI results are overwritten by x86 machines. Should be a Jenkin issue. Monitor if this corner will happen again. - Juraj
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx. Basic skeleton of docker topology done. Moving on to deploying the scripts. Juraj is able to modify, execute the scripts in Container. Things to fix is scripts for 1-link 1-node topology and interfaces binding to VPP.
      • Kernel Migration on mcbin. Juraj is able to build all the images, but got kernel panic. Try with the more recent uBoot version.
      • Juraj to investigate further work once ThunderX2 is available.
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • Development of L2 test script is under-going now.
      • Khem will get L2 work in CI firstly, and then IP4, and other test cases.
      • Stan starts to work on performance scripts with Khem. Is able to connect Taishan machines in CSIT lab.
      • The performance topology in wiki link is to update per below file.
      • https://github.com/FDio/csit/blob/master/docs/lab/Testbeds_Xeon_Skx_Arm_Atom.md
      • Stan and Khem to come up with a summary of current status and an estimate of at least upstreaming basic L2/L3 performance suites.
        • Install Ubuntu-18.04 on Huawei Taishan servers firstly, and then investigate upstreaming performance test framework to enable Aarch64
        • Lijian to verify Ubuntu-18.04 on Taishan server.
      • Stan installed latest CSIT scripts on packet generator server(x86 NEON) and Tainshan servers in FD.io lab.
      • https://gerrit.fd.io/r/#/c/16850/
      • Some of L2 and L3 test cases passed.
  • VPP
    • Vectorization
      • [Lijian] Macro benchmarking on ThunderX2/Centriq(4%)/Taishan D05(10%) is done, data is updated into Jira. Code is in internal review.
      • [Sirshak] on ethernet-input node, investigate vectorized buffer index.
    • Memory Ordering
      • [Sirshak] svm_fifo lockless alternate algorithm for SPSC. - In internal review.
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. - Bootstrap script common.sh working and checking the test cases.
    • [Juraj] IP4 reassembly and GBP failures are fixed. Patches to enable them two are merged. No any test cases in blacklist for Aarch64 machine.
    • [Juraj] Kernel Migration on mcbin. Juraj is able to build all the images.
  • Action Items - Next Week
    • [Sirshak] - To update patch list in VPP/Aarch64 wiki

01/15/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Lijian Zhang
    • Stanislav Chlebec
    • Khemendra Kumar
    • Tina Tsou
    • Andy Wang
    • Garcia
    • Manuel
    • Gorka
    • Fede
    • Honnappa Nagarahalli
    • John Ddigilio
  • General
  • VPP Hoststack
    • iperf3 performance with Hoststack.
      • Investigating lock-less fifo and memory reordering for VPP hoststack - Sirshak
    • Sirshak is trying to set up one fd.io setup for everyone debugging VPP hoststack.
    • Gorka is trying some optimal configs for VCL. - No Updates.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • OcteonTX is received in ARM lab. Will boot it up firstly and then start doing profiling with it.
  • FD.io lab
    • ThunderX2 -
      • New Arista switch is available now. Gathering details that required by LF lab before sending the switch to CSIT lab. - Juraj
      • Cable type is confirmed. Procurement is in the process.
  • CSIT
    • VPP Path
      • IP4 reassembly and GBP failures are fixed. Patches to enable them two are merged. No any test cases in blacklist for Aarch64 machine.
      • We have voting verify on bionic. Upload nexus disabled but merge job working. - Juraj created LF ticket for nexus upload. Both master merge job and verifying job are working fine.
      • ARM CI results are overwritten by x86 machines. Should be a Jenkin issue. Monitor if this corner will happen again. - Juraj
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx. Basic skeleton of docker topology done. Moving on to deploying the scripts.
      • Kernel Migration on mcbin. Juraj is able to build all the images.
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • Development of L2 test script is under-going now.
      • Khem will get L2 work in CI firstly, and then IP4, and other test cases.
      • Stan starts to work on performance scripts with Khem. Is able to connect Taishan machines in CSIT lab.
      • The performance topology in wiki link is to update per below file.
      • https://github.com/FDio/csit/blob/master/docs/lab/Testbeds_Xeon_Skx_Arm_Atom.md
      • Stan and Khem to come up with a summary of current status and an estimate of at least upstreaming basic L2/L3 performance suites.
  • VPP
    • Vectorization
      • [Lijian] Macro benchmarking on ThunderX2/Centriq(4%)/Taishan D05(10%) is done, data is updated into Jira. Code is in internal review.
    • Memory Ordering
      • [Sirshak] svm_fifo lockless alternate algorithm for SPSC. - In internal review.
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. - Bootstrap script common.sh working and checking the test cases.
    • [Juraj] IP4 reassembly and GBP failures are fixed. Patches to enable them two are merged. No any test cases in blacklist for Aarch64 machine.
    • [Juraj] Kernel Migration on mcbin. Juraj is able to build all the images.
  • Action Items - Next Week
    • [Sirshak] - To update patch list in VPP/Aarch64 wiki

01/08/2019

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Lijian Zhang
    • Stanislav Chlebec
    • Khemendra Kumar
  • General
  • VPP Hoststack
    • iperf3 performance with Hoststack.
      • Investigating lock-less fifo and memory reordering for VPP hoststack - Sirshak
    • Sirshak is trying to set up one fd.io setup for everyone debugging VPP hoststack.
    • Gorka is trying some optimal configs for VCL. - No Updates.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. - Bootstrap script common.sh working and checking the test cases.
    • [Lijian] Working on IP4 reassembly and GBP failures. - fixed. Juraj has upstreamed patched to enable these two tests.
    • [Sirshak] Kernel Migration mcbin. Juraj is working on based on Jianlin's suggestion.
    • [Andy] Getting a new Arista switch next year.
  • VPP
    • Vectorization
      • [Lijian] working on vectorized memory copy - Macro benchmarking is done and data is updated to Jira.
    • Memory Ordering
      • [Sirshak] svm_fifo lockless alternate algorithm for SPSC. - In internal review.
  • CSIT
    • VPP Path
  • VPP Path Failures
      • We have voting verify on bionic. Upload nexus disabled but merge job working. - Juraj created LF ticket for nexus upload. Both merge job and verifying job are working fine.
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx. Basic skeleton of docker topology done. Moving on to deploying the scripts.
      • thunderx2: Juraj working with LF to get this resolved.
      • mcbin: Juraj can contact Jianlin if needed.
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • Development of L2 test script is under-going now.
      • Khem will get L2 work in CI firstly, and then IP4, and other test cases.
      • Stan is starting working on VPP performance test. Khem to send email to Stan on VPP performance testing stuff.
  • FD.io lab
    • New Arista switch to be proccured next year.
    • ThunderX2 - Racked. Andy is trying to buy cables compatible to Intel XL710. Juraj to confirm info required by lab people before sending out the cables.
  • Action Items - Next Week

12/18/2018

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Tina Tsou
    • Stanislav Chlebec
    • Avinash
    • Khemendra
  • General
  • VPP Hoststack
    • iperf3 performance with Hoststack.
      • Investigating lock-less fifo and memory reordering for VPP hoststack - Sirshak
    • Sirshak is trying to set up one fd.io setup for everyone debugging VPP hoststack.
    • Gorka is trying some optimal configs for VCL. - No Updates.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. - Bootstrap script common.sh working.
    • [Lijian] Working on IP4 reassembly and GBP failures. - Some preliminary on gbp waiting Neale. Juraj to give access to Lijian to investigate on ThunderX.
    • [Sirshak] Kernel Migration mcbin. Status: Jianlin to work with Juraj to get fd.io mcbins up and running. Sirshak to setup a meeting.
    • [Andy] Getting a new Arista switch next year.
  • VPP
    • Vectorization
      • [Lijian] working on vectorized memory copy - Still benchmarking and setting it up for internal review.
      • [Lijian] Patch for compiling issue with GCC-8.x is under community review. Status: No updtaes.
      • [Lijian] Patch for fixing StringTest failure is under community review. Status: Abandoned.
      • [Lijian] Patch for CDP failure is under community review. Status: No updates.
    • Memory Ordering
      • [Sirshak] svm_fifo lockless alternate algorithm for SPSC.
  • CSIT
    • VPP Path
  • VPP Path Failures
    • https://jira.fd.io/browse/VPP-1475 - IP4 random reassembly failure in master, also seen on x86
    • https://jira.fd.io/browse/VPP-1491 - GBP L3/L2 Endpoint Learning failure
      • We have voting verify on bionic. Upload nexus disabled but merge job working. Juraj to create LF ticket for nexus upload.
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx. Basic skeleton of docker topology done. Moving on to deploying the scripts.
      • thunderx2: Sirshak working with LF to get this resolved.
      • mcbin: Sirshak to setup a meeting between Juraj and Jianlin.
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • Development of L2 test script is under-going now.
      • Khem will get L2 work in CI firstly, and then IP4, and other test cases.
  • FD.io lab
    • New Arista switch to be proccured next year.
    • ThunderX2 - Racked. IPMI Static IP configuration missing. Sirshak with LF.
  • Action Items - Next Week

12/11/2018

  • Attendees
    • Sirshak Das
    • Juraj Linkeš
    • Tina Tsou
    • Stanislav Chlebec
  • VPP Hoststack
    • iperf3 performance with Hoststack. - Sirshak has done some preliminary bench-marking, and compare kernel and VPP hoststack performance. Three case: kernel to kernel; kernel to VPP hoststack; VPP hoststack to VPP hoststack
    • ongoing perf analysis. One patch(https://gerrit.fd.io/r/#/c/16184/) is merged, and the other one is under internal review.
    • Investigating lock-less fifo and memory reordering for VPP hoststack - Sirshak
    • Sirshak is trying to set up one CSIT setup for everyone debugging VPP hoststack. Will share setup info
    • Gorka is trying some optimal configs for VCL.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. Two scripts of L2 performance suites for CI management repository are done, investigating on for CSIT repository, and three more scripts to be developed.
    • [Lijian] Working on IP4 reassembly and GBP failures
    • [Sirshak] Kernel Migration mcbin. Status: Sirshak to try inputs from garcia and damjan. - no progress so far. - To confirm with Jianling and Joyce.
  • VPP
    • Vectorization
      • [Lijian] working on vectorized memory copy - Second priority, no update so far.
      • [Lijian] Patch for compiling issue with GCC-8.x is under community review.
      • [Lijian] Patch for fixing StringTest failure is under community review.
      • [Lijian] Patch for CDP failure is under community review.
    • Memory Ordering
      • [Sirshak] To start work on Arithmetic and Logic relaxed functions.
  • VPP Path failures
  • CSIT
    • VPP Path
      • Actually, everything is ready. The only thing is to get CI patch merged.
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx is in place, but there are errors. Will continue investigation.
      • thunderx2: Racked. Lack of static IP. Sirshak gave a work-around to fix lacking of static IP to Anton.
      • mcbin: Kernel issue yet to try suggestion from Garcia and Damjan. To confirm with Jianling and Joyce - Lijian
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • Development of L2 test script is under-going now. Khem will get L2 work in CI firstly, and then IP4, and other test cases.
  • FD.io lab
    • Arista switch is still not working. Andy and Anton are working on the exact requirement for the switch, and two possible option: Andy to replace the Arista or buy a new one.
    • ThunderX2 - Racked. Lack of IP.
  • Action Items - Next Week
    • [Lijian] to continue to investigate make test failures.
    • [Andy] to work with Anton to resolve Arista problem.

12/04/2018

  • Attendees
    • Sirshak Das
    • Andy Wang
    • Juraj Linkeš
    • Khemendra
    • Garcia
    • Manuel
    • Gorka
    • Fede
    • Tina Tsou
  • VPP Hoststack
    • iperf3 performance with Hoststack. - Sirshak has done some preliminary bench-marking, and compare kernel and VPP hoststack performance. Three case: kernel to kernel; kernel to VPP hoststack; VPP hoststack to VPP hoststack
    • ongoing perf analysis. Two patches ongoing. One is upstreamed and the other is under internal review. Hotpots on memory copy or maybe other stuff.
    • Sirshak is trying to set up one CSIT setup for everyone debugging VPP hoststack. Will share setup info
    • Gorka is trying some optimal configs for VCL.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. Two scripts of L2 performance suites for CI management repository are done, investigating on for CSIT repository, and three more scripts to be developed.
    • [Lijian] VPP dlmalloc crash issue root-caused and fixed by maintainer. Florin Coras fixed time-out issues.
    • [Sirshak] Kernel Migration mcbin. Status: Sirshak to try inputs from garcia and damjan. - no progress so far. - To confirm with Jianling and Joyce.
  • VPP
    • Vectorization
      • [Lijian] working on vectorized memory copy - Second priority, no update so far.
      • [Lijian] Patch for compiling issue with GCC-8.x is under internal review.
      • [Lijian] Patch for fixing StringTest failure is under internal review.
    • Memory Ordering
      • [Sirshak] To start work on Arithmetic and Logic relaxed functions.
  • CSIT
    • VPP Path
    • VPP Device
      • thunderx: 1-node topology on cavium thunderx is in place, but there are errors. Will continue investigation.
      • thunderx2: Racked. Lack of IP. To confirm with Anton.
      • mcbin: Kernel issue yet to try suggestion from Garcia and Damjan. To confirm with Jianling and Joyce - Lijian
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • Development of L2 test script is under-going now. Khem will get L2 work in CI firstly, and then IP4, and other test cases.
  • FD.io lab
    • Arista switch is still not working. Andy and Anton are working on the exact requirement for the switch, and two possible option: Andy to replace the Arista or buy a new one.
    • ThunderX2 - Racked. Lack of IP.
  • Action Items - Next Week
    • [Lijian] to continue to investigate make test failures.
    • [Andy] to work with Anton to resolve Arista problem.


11/27/2018

  • Attendees
    • Juraj Linkeš
    • Khemendra
    • Garcia
    • Manuel
    • Gorka
    • Fede
    • Tina Tsou
  • VPP Hoststack
    • iperf3 performance with Hoststack. - Sirshak has done some preliminary bench-marking, and compare kernel and VPP hoststack performance.
    • ongoing perf analysis, two patches ongoing. Hotpots on memory copy or maybe other stuff. Will share patches with community.- Sirshak
    • Sirshak is trying to set up one CSIT setup for everyone debugging VPP hoststack. Will share setup info
    • Gorka is trying some optimal configs for VCL.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • Alternate test cases.
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. Status: Working with Juraj to get scripts ready for jobs. - Need to prepare some scripts. First to understand how the script works and then add more options.
    • [Lijian] Status on VPP path failures. Status: Still debugging. Still at early study stage.
    • [Sirshak] Kernel Migration mcbin. Status: Sirshak to try inputs from garcia and damjan. - no progress so far.
  • VPP
    • Vectorization
      • [Lijian] working on vectorized memory copy
    • Memory Ordering
      • [Sirshak] To start work on Arithmetic and Logic relaxed functions.
  • CSIT
    • VPP Path
      • 3 failures currently stalling deployment.
      • VPP-1476, VPP-1475, VPP-1478
      • These failures are seen on Debian x86 VM also.
      • Parallelization(n=32) is resulting in failures. Seems also be caused by below two patches.
      • VPP-1490, caused by https://gerrit.fd.io/r/#/c/15106/ and https://gerrit.fd.io/r/#/c/15534/.
      • VPP-1491, VPP-1497 about parallazation and GBP failure are filed.
      • Get CSIT/Aarch64 pass with partial test cases - Juraj
    • VPP Device
      • thunderx: Juraj created a LF tkt for wiring the 1-node topology on cavium thunderx.
      • thunderx2: to be racked by this Friday.
      • mcbin: Kernel issue yet to try suggestion from Garcia and Damjan.
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • L2 test is working now manually. Khem is trying to get it work in CI, and then IP4, and other test cases.
  • FD.io lab
    • Arista switch is missing cable. Andy will send tracking no. for cables.
    • ThunderX2 - to be racked by this Friday.
  • Action Items - Next Week
    • [Lijian] to investigate VPP-1490 issue.
    • [Andy] Andy will send tracking no. for cables.

11/20/2018

  • Attendees
    • Sirshak Das
    • Andy Wang
    • Juraj Linkeš
    • Khemendra
    • Garcia
    • Manuel
    • Gorka
    • Fede
    • Tina Tsou
  • VPP Hoststack
    • iperf3 performance with Hoststack. - Sirshak has done some preliminary bench-marking, and compare kernel and VPP hoststack performance.
    • ongoing perf analysis, two patches ongoing. Hotpots on memory copy or maybe other stuff. Will share patches with community.- Sirshak
    • Sirshak is trying to set up one CSIT setup for everyone debugging VPP hoststack. Will share setup info
    • Gorka is trying some optimal configs for VCL.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • Alternate test cases.
  • Action Items - Last Week
    • [Sachin] to introduce RFC for IPsec offload support in DPDK plugin.
    • [Khem] Deployment of only L2 CSIT performance suite. Status: Working with Juraj to get scripts ready for jobs. - Need to prepare some scripts. First to understand how the script works and then add more options.
    • [Lijian] Status on VPP path failures. Status: Still debugging. Still at early study stage.
    • [Sirshak] Kernel Migration mcbin. Status: Sirshak to try inputs from garcia and damjan. - no progress so far.
  • VPP
    • Vectorization
      • [Lijian] working on vectorized memory copy
    • Memory Ordering
      • [Sirshak] To start work on Arithmetic and Logic relaxed functions.
  • CSIT
    • VPP Path
      • 3 failures currently stalling deployment.
      • VPP-1476, VPP-1475, VPP-1478
      • These failures are seen on Debian x86 VM also.
      • Parallelization(n=32) is resulting in failures. Seems also be caused by below two patches.
      • VPP-1490, caused by https://gerrit.fd.io/r/#/c/15106/ and https://gerrit.fd.io/r/#/c/15534/.
      • VPP-1491, VPP-1497 about parallazation and GBP failure are filed.
      • Get CSIT/Aarch64 pass with partial test cases - Juraj
    • VPP Device
      • thunderx: Juraj created a LF tkt for wiring the 1-node topology on cavium thunderx.
      • thunderx2: to be racked by this Friday.
      • mcbin: Kernel issue yet to try suggestion from Garcia and Damjan.
    • VPP Performance Test
      • Working ongoing on writing scripts for Performance Jobs.
      • L2 test is working now manually. Khem is trying to get it work in CI, and then IP4, and other test cases.
  • FD.io lab
    • Arista switch is missing cable. Andy will send tracking no. for cables.
    • ThunderX2 - to be racked by this Friday.
  • Action Items - Next Week
    • [Lijian] to investigate VPP-1490 issue.
    • [Andy] Andy will send tracking no. for cables.


11/12/2018

  • Attendees
    • Sirshak Das
    • Andy Wang
    • Juraj Linkeš
    • Khemendra
    • Garcia
    • Gorka
  • VPP Hoststack
    • iperf3 performance with Hoststack. - Sirshak has done some preliminary bench-marking, and compare kernel and VPP hoststack performance.
    • ongoing perf analysis, two patches ongoing. Hotpots on memory copy or maybe other stuff. - Sirshak
    • Sirshak is trying to set up one CSIT setup for everyone debugging VPP hoststack. Will share setup info
    • Gorka is trying some optimal configs for VCL.
    • VPP on both sides(iperf3 server and client) give a boost.(Reason unknown).
    • Alternate test cases.
    • khem to get more information on benchmarking DMM. Khem to send the information to

Status Report Ligato/Contiv

Capture LandC.PNG

Retrieved from "https://wiki.fd.io/index.php?title=VPP/AArch64&oldid=11513"