[
https://issues.apache.org/jira/browse/HBASE-8450?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13656219#comment-13656219
]
Jean-Daniel Cryans commented on HBASE-8450:
-------------------------------------------
{quote}
The bigger block size, the lesser namenode rpc call.
We have a cluster with hbase.regionserver.hlog.blocksize=128m and the write
performance is improved a lot compare with
hbase.regionserver.hlog.blocksize=64m, the cell is big and most is 15K+.
{quote}
I don't think less RPCs to the NN really matters a lot here, other parts of
HBase abuse the NN a lot more, but most likely your write performance improved
because you are spending less time locking when rolling the logs. I pushed this
to the extreme and tried 1GB and 2GB blocksizes and that's when I got the bad
performance mostly.
Now that I think about it more, it might make sense to keep bigger logs but
keep the block size itself low. For example, have
hbase.regionserver.hlog.blocksize=64MB but
hbase.regionserver.logroll.multiplier=3.95 instead of 0.95. Is this something
you have tried or considered [~aoxiang]?
> Update hbase-default.xml and general recommendations to better suit current
> hw, h2, experience, etc.
> ----------------------------------------------------------------------------------------------------
>
> Key: HBASE-8450
> URL: https://issues.apache.org/jira/browse/HBASE-8450
> Project: HBase
> Issue Type: Task
> Components: Usability
> Reporter: stack
> Priority: Critical
> Fix For: 0.95.1
>
> Attachments: 8450.txt
>
>
> This is a critical task we need to do before we release; review our defaults.
> On cursory review, there are configs in hbase-default.xml that no longer have
> matching code; there are some that should be changed because we know better
> now and others that should be amended because hardware and deploys are bigger
> than they used to be.
> We could also move stuff around so that the must-edit stuff is near the top
> (zk quorum config. is mid-way down the page) and beef up the descriptions --
> especially since these descriptions shine through in the refguide.
> Lastly, I notice that our tgz does not "include" an hbase-default.xml other
> than the one bundled up in the jar. Maybe we should make it more accessible.
--
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
