VPP Performance Changes 3n-skx 2t1c pdr

  • Test Case: Test cases executed for: PDR throughput on 3n-skx with xxv710.

  • YYMM.V-E: [Mpps +- StDev] results for VPP (YY year, MM month, V minor version) tested in CSIT (E environment version, “2” for CSIT rls2001, “4” for CSIT rls2005).

  • YYMM.V-E vs YYMM.V-E: [% +- StDev] Relative difference, latter is the base. [N] for Root Cause Analysis (RCA) reference, see RCA list below.

  • 2001.0-4 vs 2001.0-2: Compares CSIT environment current vs previous with previous VPP code.

  • 2005.0-4 vs 2001.0-4: Compares VPP code current vs previous in the current CSIT environment

Root Cause Analysis:

  • [4] DONE Progression in high performance tests due to Gerrit 26445. TRex setting improvement. [CSIT-1740]

  • [5] DONE Progression in tests with VPP (l2xc) inside a VM due to Gerrit 25309. Make VPP in VM worker placement deterministic. [CSIT-1738]

  • [6] DONE Apparent regression in CSIT is caused by VPP renaming the most performant IPsec plugin, thus the current CSIT version used with the previous VPP version measures the less performant VPP IPsec plugin. [CSIT-1742]

  • [7] POSTPONED Possible anomaly, but stdev of results is high. Performance differences considered to probably be just noise.

  • [8] DONE Regressions in some tests with VPP in VM due to Gerrit 26744. Switch to clang-9 compiler identified as cause. Tests using l2xcbase as vswitch with small stdev used for bisect. [CSIT-1741]

  • [9] OPEN Progression in all ip6 scale tests. Changes to vppinfra bihash may have improved code efficiency. [CSIT-1735]

  • [10] OPEN Progression in ip6 base tests with dot1q. [CSIT-1746]

  • [11] OPEN Progression in memif tests, mainly in avf-eth-l2xcbase-eth-2memif-1dcr. Two progression candidates found: Gerrit 24053 and Gerrit 26744. Performance results are not consistent enough to separate the two from other possible causes. [CSIT-1744]

  • [12] OPEN Progression in some l2bdscale tests. [CSIT-1745]

  • [16] OPEN Apparent progression in IPsec tests, only for aes128cbc-hmac512sha. Considering [6], the positive difference is not large enough, so there was probably a VPP regression. [CSIT-1748]

  • [17] OPEN Progression in some vhost tests (dpdk in VM). [CSIT-1749]