You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@trafficserver.apache.org by "Luca Bruno (JIRA)" <ji...@apache.org> on 2016/02/04 18:23:39 UTC
[jira] [Updated] (TS-4173) High latency with
--enable-linux-native-aio
[ https://issues.apache.org/jira/browse/TS-4173?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Luca Bruno updated TS-4173:
---------------------------
Description:
Problem:
My average disk latency on rotational disk with 10000RPM is around 5ms.
However ATS response times, regardless of having low or high concurrent requests, is on average 20ms.
The cause:
With --enable-linux-native-aio, the AIO period time is hardcoded to 10ms (AIO_PERIOD in AIO.cc).
That's really suboptimal with disk latency that is lower than 10ms.
Without --enable-linux-native-aio I get on average 5ms response times which resemble the disk read latency.
Possible solution:
Instead of waiting for 10ms unconditionally:
- Decide (don't know how yet) a number of disk ops NOPS.
- If the ops queue size >= NOPS, trigger mainAIOEvent, then reset the timer again to AIO_PERIOD.
was:
Problem:
My average disk latency on rotational disk with 10000RPM is however around 5ms.
However ATS response times, regardless of having low or high concurrent requests, is on average 20ms.
The cause:
With --enable-linux-native-aio, the AIO period time is hardcoded to 10ms (AIO_PERIOD in AIO.cc).
That's really suboptimal with disk latency that is lower than 10ms.
Without --enable-linux-native-aio I get on average 5ms response times which resemble the disk read latency.
Possible solution:
Instead of waiting for 10ms unconditionally:
- Decide (don't know how yet) a number of disk ops NOPS.
- If the ops queue size >= NOPS, trigger mainAIOEvent, then reset the timer again to AIO_PERIOD.
> High latency with --enable-linux-native-aio
> -------------------------------------------
>
> Key: TS-4173
> URL: https://issues.apache.org/jira/browse/TS-4173
> Project: Traffic Server
> Issue Type: Bug
> Components: Cache
> Reporter: Luca Bruno
>
> Problem:
> My average disk latency on rotational disk with 10000RPM is around 5ms.
> However ATS response times, regardless of having low or high concurrent requests, is on average 20ms.
> The cause:
> With --enable-linux-native-aio, the AIO period time is hardcoded to 10ms (AIO_PERIOD in AIO.cc).
> That's really suboptimal with disk latency that is lower than 10ms.
> Without --enable-linux-native-aio I get on average 5ms response times which resemble the disk read latency.
> Possible solution:
> Instead of waiting for 10ms unconditionally:
> - Decide (don't know how yet) a number of disk ops NOPS.
> - If the ops queue size >= NOPS, trigger mainAIOEvent, then reset the timer again to AIO_PERIOD.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)