hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ramkrishna.s.vasudevan (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-17644) Create ByteBuffered cells only while creating a cell from MSLAB
Date Tue, 14 Feb 2017 06:18:42 GMT

    [ https://issues.apache.org/jira/browse/HBASE-17644?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15865150#comment-15865150
] 

ramkrishna.s.vasudevan commented on HBASE-17644:
------------------------------------------------

bq.That there is a mix of types makes a difference? Or is it just that we always do millions
of comparisons and offheap comparisons are slower? 
Our tests with PE shows that it is the mix that creates differences. Offheap ByteBuffered
based comparisons are infact faster.
But am trying to prove this with JMH but am  not getting the same convincing result. But PE
results are clear.

> Create ByteBuffered cells only while creating a cell from MSLAB
> ---------------------------------------------------------------
>
>                 Key: HBASE-17644
>                 URL: https://issues.apache.org/jira/browse/HBASE-17644
>             Project: HBase
>          Issue Type: Sub-task
>          Components: regionserver
>    Affects Versions: 2.0.0
>            Reporter: ramkrishna.s.vasudevan
>             Fix For: 2.0.0
>
>
> We create a cell out of the Bytebuffer that is returned by MSLAB. This BB can be offheap
or onheap type. If MSLAB is full we return onheap byte buffer. So the place where we create
a cell by copying data to this buffer we either create KeyValue or OffheapKV based on the
buffer type. So what we saw in tests is that since we have a combination of Cells the comparisons
that happens when adding to memstore happens millions of times and that has an impact on the
performance of write path. In read path this is not significant enougth (though we have plans
to just create one type of cells every where in Server side). 



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Mime
View raw message