[
https://issues.apache.org/jira/browse/ACCUMULO-4114?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15109248#comment-15109248
]
Eric Newton commented on ACCUMULO-4114:
---------------------------------------
Current theory is that the metadata table is being squeezed into a smaller
space, while needing to hold more and more data.
I have attached another graph. This time I simply counted the number of MinC's
and MajC's of the metadata table that have taken place during the test run. It
certainly correlates with lower performance.
> examine performance decrease with many tablets
> ----------------------------------------------
>
> Key: ACCUMULO-4114
> URL: https://issues.apache.org/jira/browse/ACCUMULO-4114
> Project: Accumulo
> Issue Type: Task
> Components: client, tserver
> Reporter: Eric Newton
> Assignee: Eric Newton
> Fix For: 1.8.0
>
> Attachments: MetadataCompactionsPerTest.png, PerfWithMoreTablets.png
>
>
> While writing a performance test for ACCUMULO-4112, I noticed the performance
> drop for large numbers of tablets hosted on a single server.
> Some decrease in performance is expected since the various queues and buffers
> are being divided among a greater number of tablets, but a 5x decrease is
> unexpected. Given other fixes (ACCUMULO-3423, for example) I would not
> expect performance to drop off so dramatically.
> Investigate the performance of conditional updates going to many tablets, vs
> one tablet.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
