You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Andrey Gura (JIRA)" <ji...@apache.org> on 2018/04/04 18:17:00 UTC

[jira] [Updated] (IGNITE-7820) Investigate and fix perfromance drop of WAL for FSYNC mode

     [ https://issues.apache.org/jira/browse/IGNITE-7820?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Andrey Gura updated IGNITE-7820:
--------------------------------
    Fix Version/s:     (was: 2.5)
                   2.6

> Investigate and fix perfromance drop of WAL for FSYNC mode
> ----------------------------------------------------------
>
>                 Key: IGNITE-7820
>                 URL: https://issues.apache.org/jira/browse/IGNITE-7820
>             Project: Ignite
>          Issue Type: Bug
>    Affects Versions: 2.4
>            Reporter: Andrey Gura
>            Assignee: Andrey Gura
>            Priority: Major
>             Fix For: 2.6
>
>
> WAL performance drop was introduced by https://issues.apache.org/jira/browse/IGNITE-6339 fix. In order to provide better performance for {{FSYNC}} WAL mode {{FsyncModeFileWriteAheadLogManager}} implementation was added as result of fix issue https://issues.apache.org/jira/browse/IGNITE-7594.
> *What we know about this performance drop:*
> * It affects {{IgnitePutAllBenchmark}} and {{IgnitePutAllTxBenchmark}} and measurements show 10-15% drop and ~50% drop accordingly.
> * It is reproducible not for all hardware configuration. That is for some configuration we see performance improvements instead of drop.
> * It is reproducible for [Many clients --> One server] topology.
> * If {{IGNITE_WAL_MMAP == false}} then we have better performance.
> * If {{fsyncDelay == 0}} then we have better performance.
> *What were tried during initial investigation:*
> * Replacing of {{LockSupport.park/unpark}} to spin leads to improvement about 2%.
> * Using {{FileWriteHandle.fsync(null)}} (unconditional flush) instead of {{FileWriteHandle.fsync(position)}} (conditional flush) doesn't affect benchmarks.
> *What should we do:*
> Investigate the problem and provide fix or recommendation for system tuning.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)