Heng Chen commented on HBASE-16652:

I add time log about 'get' operation, and do test with pe command,  3 runs.
bin/hbase org.apache.hadoop.hbase.PerformanceEvaluation --rows=10000 --nomapred 
increment/append 10

calculate the AVG time about 'get' action in increment/append,  the results 
show below:

Increment:  55693ns,  57180ns,  56374ns
Append:  50174ns,  50101ns,  52870ns

'Get' in increment is about 10% slower than append.  It matches the profile i 
have uploaded.   And the profile shows that 
{{org.apache.hadoop.hbase.regionserver.StoreScanner.seekAsDirection(Cell)}} is 
5% in 'Increment' profile,  but this action NOT appear in 'Append'.  Upload the 
relates picture.

> Figure out performance difference between increment and append
> --------------------------------------------------------------
>                 Key: HBASE-16652
>                 URL: https://issues.apache.org/jira/browse/HBASE-16652
>             Project: HBase
>          Issue Type: Sub-task
>            Reporter: Heng Chen
>              Labels: performance
>         Attachments: append.jfr, increment.jfr
> When do performance test in HBASE-16625,  i found it has the very big 
> difference between Append and Increment (append is about 37% faster than 
> increment).
> As [~stack] mentioned in 
> https://issues.apache.org/jira/browse/HBASE-16610?focusedCommentId=15493166&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15493166,
>    append and increment has been unified in server-side,  and they looks the 
> same in client-side. 
> This issue is to figure out why the performance looks different between them.

This message was sent by Atlassian JIRA

Reply via email to