[
https://issues.apache.org/jira/browse/TS-2894?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14353916#comment-14353916
]
Sudheer Vinukonda commented on TS-2894:
---------------------------------------
Hi Leif - We have an internal patch running in prod (although, I am not sure if
it was ever needed to be activated, since, the flapping issues didn't occur,
following a h/w upgrade). If someone can review the patch and provide feedback
on whether this is needed upstream, I can commit/push it.
> Spdy slow start..
> -----------------
>
> Key: TS-2894
> URL: https://issues.apache.org/jira/browse/TS-2894
> Project: Traffic Server
> Issue Type: Improvement
> Components: SPDY
> Reporter: Sudheer Vinukonda
> Assignee: Sudheer Vinukonda
> Labels: yahoo
> Fix For: 5.3.0
>
> Attachments: TS-2894.diff
>
>
> When production testing with spdy/5.0.0, we ran into an issue in some of our
> systems, where, the spdy hosts would flap constantly due to the flood of
> requests. We further noticed that, where the 4.0.x version or 5.0.0 w/ spdy
> turned off, would recover quickly following a restart, spdy enabled hosts
> would continue to receive flood of requests and continue to flap. During this
> time, traffic server is generally busy reading from the disk and can not
> handle too many requests, and is made miserable by spdy's support of multiple
> concurrent streams.
> To handle such a sudden flood of requests, I'm implementing a simple slow
> start mechanism with spdy. The idea is to increase the
> max_concurrent_streams_in gradually based on a configured timer, rather than
> use the configured value right away. The steps I chose to implement are 1,
> 25, 50, 75 and 100% of the configured max_concurrent_streams_in. Note that,
> currently,
> max_concurrent_streams_in only affects new spdy sessions. Existing sessions
> (if any) would continue to use their older values.
> Not too sure, if everyone would be interested in this..but, thought of still
> uploading my patch, incase, someone is interested.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)