More info after investigating further: the issue seems related to the fact that the RDMA plugin is available in 19.08, which did not exist in 19.01. As a result, we no longer need the "make dpdk-install-dev DPDK_MLX5_PMD=y DPDK_MLX5_PMD_DLOPEN_DEPS=n" complication when building. The release notes for VPP 19.04 say "RDMA (ibverb) driver plugin - MLX5 with multiqueue".
For 19.01 we had configured "num-rx-queues" for each of the two interfaces used, in the dpdk dev part of the startup.conf file. After testing different choices for that it turns out that if we set "num-rx- queues 1" for each interface, then 19.01 gets the same performance problem that we see for 19.08 (i.e. only threads 0 and 1 can be used efficiently for NAT). So it appears that the reason why our 19.01 installation can use more NAT threads is that we have set larger "num- rx-queues" values. For 19.08 however, the "num-rx-queues" values seem to be ignored, presumably because the RDMA plugin is used. Is it correct that the dpdk dev num-rx-queues option is ignored when the RDMA plugin is used? How can we add more queues or polling threads to RDMA interfaces so that we can use more NAT workers? Best regards, Elias On Thu, 2019-10-03 at 07:28 +0000, Elias Rudberg wrote: > As we are about to switch from VPP 19.01 to 19.08 we encountered a > problem with NAT performance. We try to use the same settings (as far > as possible) for 19.08 as we did for 19.01, on the same computer. > > [...]
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#14105): https://lists.fd.io/g/vpp-dev/message/14105 Mute This Topic: https://lists.fd.io/mt/34379814/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
