[
https://issues.apache.org/jira/browse/TS-880?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13080170#comment-13080170
]
William Bardwell commented on TS-880:
-------------------------------------
I have a new fix based on making UnixNetVConnection::reenable be able to wake
up another thread even when it can't get the NetHandler lock. Performance with
the fix still drops by half when the connection is on the wrong thread, but
things don't crash.
> Major performance problem with second request on same keep-alive connection
> ---------------------------------------------------------------------------
>
> Key: TS-880
> URL: https://issues.apache.org/jira/browse/TS-880
> Project: Traffic Server
> Issue Type: Bug
> Components: HTTP
> Affects Versions: 3.0.0, 2.1.4
> Environment: 32-bit TS on linux, using epoll. With
> proxy.config.http.keep_alive_enabled_out = 1
> Reporter: William Bardwell
> Assignee: William Bardwell
> Fix For: 3.1.0
>
> Attachments: performance_2nd_req.diff
>
>
> If I load the same URL through TS twice in a row to a server with keep-alives
> turned on I get really slow performance on the second request.
> E.g. (loading 212M over loopback with uncachable content, but results are
> similar with cachable content):
> % Total % Received % Xferd Average Speed Time Time Time
> Current
> Dload Upload Total Spent Left Speed
> 100 212M 100 212M 0 0 140M 0 0:00:01 0:00:01 --:--:--
> 142M*
> % Total % Received % Xferd Average Speed Time Time Time
> Current
> Dload Upload Total Spent Left Speed
> 100 212M 100 212M 0 0 6397k 0 0:00:34 0:00:34 --:--:--
> 6400k*
> If I turn off proxy.config.http.keep_alive_enabled_out the problem goes away,
> it can also be partially solved by making proxy.config.io.max_buffer_size
> really big (but it needs to be bigger for larger content, and that supports
> the comment below, and proves that it isn't the origin server's fault.)
> When I turn on debug messages it looks like the IO loop for the second
> request only wakes up every 10ms (when an epoll_wait times out) and then it
> reads and writes, for the first request it goes much faster without those
> pauses. My theory is that the issue is related to the the fact that the
> outgoing connection fd gets added to the epoll fd for the thread handling the
> first request and the first user agent fd is added to the same epoll fd, and
> it stays on that epoll fd. But the new user agent request is on a different
> epoll fd which I assume means is tied to a different thread.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
