From a8c3f441fa595311f60bcc634a2720f204ced733 Mon Sep 17 00:00:00 2001 From: Maciek Konstantynowicz Date: Thu, 2 May 2019 19:20:38 +0100 Subject: report 1904: updated methodology (nfv density, startup.conf) and vpp perf rls notes Change-Id: If6d24511d5d29dcc25c21437364e0da15af3cca1 Signed-off-by: Maciek Konstantynowicz --- .../methodology_vpp_startup_settings.rst | 31 ++++++++++++---------- 1 file changed, 17 insertions(+), 14 deletions(-) (limited to 'docs/report/introduction/methodology_vpp_startup_settings.rst') diff --git a/docs/report/introduction/methodology_vpp_startup_settings.rst b/docs/report/introduction/methodology_vpp_startup_settings.rst index 81c5e4e73e..e3e8d29b23 100644 --- a/docs/report/introduction/methodology_vpp_startup_settings.rst +++ b/docs/report/introduction/methodology_vpp_startup_settings.rst @@ -1,29 +1,32 @@ VPP Startup Settings -------------------- -CSIT code manipulates a number of VPP settings in startup.conf for optimized -performance. List of common settings applied to all tests and test -dependent settings follows. +CSIT code manipulates a number of VPP settings in startup.conf for +optimized performance. List of common settings applied to all tests and +test dependent settings follows. -See `VPP startup.conf`_ -for a complete set and description of listed settings. +See `VPP startup.conf`_ for a complete set and description of listed +settings. Common Settings ~~~~~~~~~~~~~~~ -List of vpp startup.conf settings applied to all tests: +List of VPP startup.conf settings applied to all tests: #. heap-size - set separately for ip4, ip6, stats, main depending on scale tested. -#. no-tx-checksum-offload - disables UDP / TCP TX checksum offload in DPDK. - Typically needed for use faster vector PMDs (together with +#. no-tx-checksum-offload - disables UDP / TCP TX checksum offload in + DPDK. Typically needed for use faster vector PMDs (together with no-multi-seg). -#. buffers-per-numa - increases number of buffers allocated, needed - in scenarios with large number of interfaces and worker threads. - Value is per CPU socket. Default is 16384. CSIT is setting statically - 107520 buffers per CPU thread (215040 if HTT is enabled). This value is also - maximum possible amount limited by number of memory mappings in DPDK - libraries for 2MB Hugepages used in CSIT. +#. buffers-per-numa - sets a number of memory buffers allocated + to VPP per CPU socket. VPP default is 16384. Needs to be increased for + scenarios with large number of interfaces and worker threads. To + accommodate for scale tests, CSIT is setting it to the maximum possible + value corresponding to the limit of DPDK memory mappings (currently + 256). For Xeon Skylake platforms configured with 2MB hugepages and VPP + data-size and buffer-size defaults (2048B and 2496B respectively), this + results in value of 215040 (256 * 840 = 215040, 840 * 2496B buffers fit + in 2MB hugepage ). For Xeon Haswell nodes value of 107520 is used. Per Test Settings ~~~~~~~~~~~~~~~~~ -- cgit 1.2.3-korg