[jira] [Commented] (HBASE-17017) Remove the current per-region latency histogram metrics

["Enis Soztutar (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-17017?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15637925#comment-15637925 ] 

Enis Soztutar commented on HBASE-17017:
---------------------------------------

I've run PE with 1000 regions in a single server:
{code}
bin/hbase pe --latency --nomapred --presplit=1000 --valueSize=1000 --rows=100000 sequentialWrite 30
{code}

We are allocating ~1M LongAdder (former Counter) objects which is crazy. With a simple patch, the allocations goes down to less than 0.5% of heap so that JFR does not show it. The runtime for PE improves 17% because we are not spending time on this code path any more:
{code}
    private LongAdder[] createCounters(int numBins) {
      return Stream.generate(LongAdder::new).limit(numBins + 3).toArray(LongAdder[]::new);
    }
{code}

See attached screenshots. 

> Remove the current per-region latency histogram metrics
> -------------------------------------------------------
>
>                 Key: HBASE-17017
>                 URL: https://issues.apache.org/jira/browse/HBASE-17017
>             Project: HBase
>          Issue Type: Sub-task
>          Components: metrics
>    Affects Versions: 2.0.0, 1.4.0
>            Reporter: Duo Zhang
>            Assignee: Enis Soztutar
>             Fix For: 2.0.0, 1.4.0
>
>




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