[jira] [Comment Edited] (HBASE-7755) Experiment with LAB in BlockEndcoding

["Lars Hofhansl (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-7755?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13571029#comment-13571029 ] 

Lars Hofhansl edited comment on HBASE-7755 at 2/5/13 4:30 AM:
--------------------------------------------------------------

After thinking about this a bit more... It does not make sense to have the chunk size larger than an individual block's size, and thus it makes sense to create a new LAB when a block is scanned, and to do that for each block. The LAB is almost free to allocate; and since this is done only when we seek+read (i.e. scanning) and only when block encoding is enabled we'll be copying a lot of bytes during the decoding anyway.

>From that viewpoint BufferedEncodedSeeker is in fact the right place. The only part missing is the config option to disable this (which turns out to be a bit more tricky to do nicely - without passing information down a 10-depth call stack).
Also the alloc size should probably be less than chunksize. I'm thinking to make 1/10th of the chunk size.
                
      was (Author: lhofhansl):
    After thinking about this a bit more... It does not make sense to the chunk size larger than an individual block's size, and thus it makes to create a new LAB when a block is scanned, and to do that for each block. The LAB is almost free allocate; and since this is done only when we seek+read (i.e. scanning) and only when block encoding is enabled we'll be copying a lot of bytes during the decoding.

>From that viewpoint BufferedEncodedSeeker is in fact the right place. The only part missing is the config option to disable this (which turns out to be a bit more tricky to do nicely - without passing information down a 10-depth call stack).
                  
> Experiment with LAB in BlockEndcoding
> -------------------------------------
>
>                 Key: HBASE-7755
>                 URL: https://issues.apache.org/jira/browse/HBASE-7755
>             Project: HBase
>          Issue Type: Bug
>            Reporter: Lars Hofhansl
>             Fix For: 0.94.6
>
>         Attachments: 7755-0.94-W_I_P_v1.txt, 7755-0.94-WORK_IN_PROGRESS.txt
>
>
> I was looking at and profiling the BlockEncoding code to figure out how to make it faster. One issue that jumped out was we call ByteBuffer.allocate(...) for each single KV.
> As an experiment I tried using the MemStoreLAB code to allocate those buffers.
> Here are some preliminary numbers, all scanning 10m rows (all in cache):
> * no encoding: 5.2s
> * FAST_DIFF without patch: 7.3s
> * FAST_DIFF with patch and small LAB: 4.1s
> * FAST_DIFF with patch and large LAB: 11s
> So this is very sensitive to the right sizing of the LAB.
> Need to do a bit more testing, but it seems that there is a chance to actually make scanning with block encoding faster than without!

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira