[
https://issues.apache.org/jira/browse/TS-1427?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
B Wyatt resolved TS-1427.
-------------------------
Resolution: Fixed
git commit: c8392efb7c10c109592ceae980c0ed6a691899d5
> PluginVC's inactivity timeout support is expensive due to Mutex and Thread
> Scheduling load
> ------------------------------------------------------------------------------------------
>
> Key: TS-1427
> URL: https://issues.apache.org/jira/browse/TS-1427
> Project: Traffic Server
> Issue Type: Bug
> Components: HTTP, TS API
> Affects Versions: 3.2.0
> Environment: 3.2 with plugin that makes use of 1 or more PluginVC's
> with long/large requests
> Reporter: B Wyatt
> Assignee: B Wyatt
> Priority: Minor
>
> PluginVC's method for handling inactivity timer is to cancel and reschedule
> an event every time activity happens. When data arrives in rapid and small
> packets this results in a few hundred cancel and resets per second per
> PluginVC. Doing so spikes futex and thread scheduling loads in the kernel.
> By contrast, the socket-based VC's use an inactivity cop that fires once a
> second to see if any connection has been inactive for too long. This only
> requires maintaining an timeout timestamp per connection which is far
> cheaper. However, it reduces the accuracy of the inactivity timeout to be
> ~1sec resolution.
> Well worth the accuracy trade to get performance back up to par in my opinion.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
