[
https://issues.apache.org/jira/browse/HBASE-11610?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14109223#comment-14109223
]
Virag Kothari commented on HBASE-11610:
---------------------------------------
bq. It's great that v3 is both simpler and more efficient!
Its ~5% faster than v2 which creates a seperate HTable for each put. Compared
to the threadLocal HTable (original patch), it is ~5% slower. (this experiments
are done over series of 5 runs).
bq. is MultiHConnection that useful? I see its size is defaulted to 1? have you
compare the performances with greater values?
All the above experiments are done with 10 HConnections. It improves the write
throughput on META server by ~15% compared to single HConnection when doing
assignment for 1M regions. But most users will have significantly lower than 1M
regions, so by default we only will have one HConnection.
I had tried with 15 connections but didn't see any increase in throughput.
bq. Each Connection will come with a pool of 256 threads (there is HBASE-11590
to improve this a little), plus the one of MultiHConnection
I believe that pool is only created when you do a getTable() on HConnection. As
we are not going the HTable route, all threads accessing regionStateStore will
use the single pool managed by MultiHConnection. We can get rid of this pool
when we have a new processBatchCallBack API which maintains a pool (though that
will be per connection).
bq. we should have some thing like hbase.regionstatestore.meta.threads.max
instead of hbase.hconnection.threads.max)
how about hbase.multihconnection.threads.max as the pool is part of
MultiHconnection?
bq. I like the patch
Thanks!
Will address comments shortly.
> Enhance remote meta updates
> ---------------------------
>
> Key: HBASE-11610
> URL: https://issues.apache.org/jira/browse/HBASE-11610
> Project: HBase
> Issue Type: Sub-task
> Reporter: Jimmy Xiang
> Assignee: Virag Kothari
> Attachments: HBASE-11610.patch, HBASE-11610_2.patch,
> HBASE-11610_v3.patch
>
>
> Currently, if the meta region is on a regionserver instead of the master,
> meta update is synchronized on one HTable instance. We should be able to do
> better.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
