You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@cassandra.apache.org by "Jay Zhuang (JIRA)" <ji...@apache.org> on 2016/12/09 01:59:58 UTC

[jira] [Commented] (CASSANDRA-12676) Message coalescing regression

    [ https://issues.apache.org/jira/browse/CASSANDRA-12676?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15734002#comment-15734002 ] 

Jay Zhuang commented on CASSANDRA-12676:
----------------------------------------

We saw huge performance gain by disabling coalescing on both 2.2.5 and 3.0.10 (about 50% average latency decrease on some clusters). I would suggest to disable it even in 3.0.x.
The attachment is test result for 5 nodes cluster with RF=3, quorum write: https://issues.apache.org/jira/secure/attachment/12842470/result.html

|| coalescing || disabled || timehorizon ||
| latency mean | 2.5 ms | 3.4 ms (+36%) |
| P95 | 3.4 ms | 4.3 ms (+26%) |
| P99 | 5.3 ms | 5.8 ms (+9.4%) |


> Message coalescing regression
> -----------------------------
>
>                 Key: CASSANDRA-12676
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-12676
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: T Jake Luciani
>              Labels: docs-impacting
>             Fix For: 4.0, 3.x
>
>         Attachments: result.html
>
>
> The default in 2.2+ was to enable TIMEHORIZON message coalescing.  After reports of performance regressions after upgrading from 2.1 to 2.2/3.0 we have discovered the issue to be this default.
> We need to re-test our assumptions on this feature but in the meantime we should default back to disabled.
> Here is a performance run [with and without message coalescing|http://cstar.datastax.com/graph?command=one_job&stats=9a26b5f2-7f48-11e6-92e7-0256e416528f&metric=op_rate&operation=2_user&smoothing=1&show_aggregates=true&xmin=0&xmax=508.86&ymin=0&ymax=91223] 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)