cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Marcus Eriksson (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-11623) Compactions w/ Short Rows Spending Time in getOnDiskFilePointer
Date Wed, 11 May 2016 08:40:12 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-11623?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15279792#comment-15279792
] 

Marcus Eriksson commented on CASSANDRA-11623:
---------------------------------------------

I ran a few small benchmarks;

* write 10M keys, autocompaction disabled, then run a major compaction, in this case this
patched version was about 10% quicker (32MB/s vs 29MB/s).
* write 10M keys, autocompaction disabled, then nodetool enableautocompaction, this showed
basically no difference.

The reason for the big difference when running major compaction is probably that trunk calls
{{getOnDiskFilePointer}} 3 times per appended partition

With these results I'll commit it to trunk only, if anyone can show datamodels with a bigger
difference, I would be happy to backport to 2.2+

> Compactions w/ Short Rows Spending Time in getOnDiskFilePointer
> ---------------------------------------------------------------
>
>                 Key: CASSANDRA-11623
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-11623
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Tom Petracca
>            Assignee: Tom Petracca
>            Priority: Minor
>             Fix For: 3.x
>
>         Attachments: compactiontask_profile.png
>
>
> Been doing some performance tuning and profiling of my cassandra cluster and noticed
that compaction speeds for my tables that I know to have very short rows were going particularly
slowly.  Profiling shows a ton of time being spent in BigTableWriter.getOnDiskFilePointer(),
and attaching strace to a CompactionTask shows that a majority of time is being spent lseek
(called by getOnDiskFilePointer), and not read or write.
> Going deeper it looks like we call getOnDiskFilePointer each row (sometimes multiple
times per row) in order to see if we've reached our expected sstable size and should start
a new writer.  This is pretty unnecessary.



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

Mime
View raw message