Hi Florin,
Thanks for the clear explanation ! I do remember that I consulted with you about this procedure, and it has been improved before. But i am failed to find the patch two days ago.
Regards,
yacan
On 5/6/2022 01:12，Florin Coras<fcoras.li...@gmail.com> wrote：
Hi Yacan,Currently rpcs from first worker to main are done through session layer and are processed by main in batches. Session queue node runs on main in interrupt mode so first worker will set an interrupt when the list of pending connects goes non-empty and main will switch to polling in the rpc handler if it notices it can’t handle pending connects in one dispatch. So, the first connect might be affected by main sleeping in epoll_pwait but subsequent connects should not, assuming we get a constant stream of connects.To test that, weizhen9612 try adding to session stanza in startup.conf: session { poll-main }. That should avoid main sleeping in epoll_pwait.Eventually, we’ll get to a point where first worker will execute the connects. Part of the changes needed are in, i.e., session pools are now realloced with barrier, but more improvements are needed.Regards,FlorinOn May 5, 2022, at 5:01 AM, liuyacan <liuya...@corp.netease.com> wrote:Hi Florin, weizhen9612:I'm not sure whether rpc for connects will be executed immediately by the main thread in the current implementation, or will it wait for the epoll_pwait in linux_epoll_input_inline() to time out.Regards,yacanHi,
Now I configure main-core to 2 and corelist-workers to 0, and find that the performance has improved significantly.
<dummyfile.0.part>
When I execute the following conmand, I find that vpp have main thread only.
#show threads
<dummyfile.1.part>
What does this situation show？
Thanks.
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#21380): https://lists.fd.io/g/vpp-dev/message/21380 Mute This Topic: https://lists.fd.io/mt/90793836/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
