[jira] [Updated] (RATIS-1730) Optimize RaftLog cache eviction time

["Song Ziyang (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/RATIS-1730?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Song Ziyang updated RATIS-1730:
-------------------------------
    Summary: Optimize RaftLog cache eviction time  (was: Use Daemon thread to evict RaftLogCache)

> Optimize RaftLog cache eviction time
> ------------------------------------
>
>                 Key: RATIS-1730
>                 URL: https://issues.apache.org/jira/browse/RATIS-1730
>             Project: Ratis
>          Issue Type: Improvement
>          Components: server
>            Reporter: Song Ziyang
>            Assignee: Song Ziyang
>            Priority: Major
>         Attachments: RATIS-1730.patch, image-20221026020057385.png, new_result.png
>
>
> Using JProfiler, I found almost 50% of AppendEntriesAsync execution time is spent on checkAndEvictCache, see the attached pic. There's also a TODO[1] suggesting to use a daemon thread to perform checkAndEvictCache. 
> The time are mainly spent on getTotalCacheSize() and countCached(). These two methods will visit every segments. When there are lots of closed segments, visiting them all is indeed a very consuming operation, not to mention that they are performed on every appendEntriesAsync.
> [1][https://github.com/apache/ratis/blob/b8117f41481f5f5861566a68bb1ea4dc7c60320c/ratis-server/src/main/java/org/apache/ratis/server/raftlog/segmented/SegmentedRaftLog.java#L402] 



--
This message was sent by Atlassian Jira
(v8.20.10#820010)