[
https://issues.apache.org/jira/browse/HBASE-17049?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15673648#comment-15673648
]
Duo Zhang commented on HBASE-17049:
-----------------------------------
Delay sync is my last trick to aggregate more syncs. If it does not work, then
I have no idea...
As in the perf result posted by [~stack], it seems that AsyncFSWAL has less
instructions but cost more time. I think it means that the AsyncFSWAL can not
make use of all the cores on the machine. FSHLog does better than us. And I
think it is reasonable. We use one thread to do everything, but FSHLog has one
disruptor thread, five sync threads, one DataStreamer thread, and one Responder
thread.
{noformat}
FSHLog
4987370.861017 task-clock (msec) # 12.922 CPUs utilized
9,934,495,287,070 cycles # 1.992 GHz
3,796,677,865,651 instructions # 0.38 insns per cycle
AsyncFSWAL
4568572.588814 task-clock (msec) # 7.454 CPUs utilized
9,292,754,813,201 cycles # 2.034 GHz
3,245,931,999,369 instructions # 0.35 insns per cycle
{noformat}
So I think we should try multiwal next? With multiwal, we will have multiple
AsyncFSWAL instance thus we could make use of more cores.
What do you guys think? [~stack] [~ram_krish].
Thanks.
> Find out why AsyncFSWAL issues much more syncs than FSHLog
> ----------------------------------------------------------
>
> Key: HBASE-17049
> URL: https://issues.apache.org/jira/browse/HBASE-17049
> Project: HBase
> Issue Type: Sub-task
> Components: wal
> Affects Versions: 2.0.0
> Reporter: Duo Zhang
> Fix For: 2.0.0
>
> Attachments: delay-sync.patch
>
>
> https://issues.apache.org/jira/browse/HBASE-16890?focusedCommentId=15647590&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15647590
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
