[
https://issues.apache.org/jira/browse/HBASE-10129?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13852581#comment-13852581
]
cuijianwei commented on HBASE-10129:
------------------------------------
To evaluate the performance change, I use the test tool YCSB to put/get
10,000,000 records to a cluster containing 5 region servers, then I compare the
average latency respectively.
For cluster without the patch, I get the put/get latency:
[PUT], AverageLatency(us), 953.5947375
[GET], AverageLatency(us), 279.0727035
For cluster with the patch, I get the put/get latency:
[PUT], AverageLatency(us), 954.3327765
[GET], AverageLatency(us), 248.4747838
It seems the performance change when applying the patch is not significant.
> support real time rpc invoke latency percentile statistics for methods of
> HRegionInterface
> -------------------------------------------------------------------------------------------
>
> Key: HBASE-10129
> URL: https://issues.apache.org/jira/browse/HBASE-10129
> Project: HBase
> Issue Type: Improvement
> Components: regionserver
> Affects Versions: 0.94.14
> Reporter: cuijianwei
> Attachments: HBASE-10129-0.94-v1.patch
>
>
> It is important for applications to get latency statistics when invoking
> hbase apis. Currently, the average latency of methods in HRegionInterface
> will be computed in HBaseRpcMetrics. However, user might expect more detail
> latency statistics, such as 75% percentile latency, 95% percentile latency,
> etc. Therefore, will it be useful if we computing latency percentiles for rpc
> invoking of region server methods?
--
This message was sent by Atlassian JIRA
(v6.1.4#6159)
