[
https://issues.apache.org/jira/browse/HDDS-2351?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16958502#comment-16958502
]
Bharat Viswanadham edited comment on HDDS-2351 at 10/24/19 4:24 AM:
--------------------------------------------------------------------
Hi [~rajesh.balamohan]
What is the test you are running? And also can we try with increasing handler
count which increase handler threads on OM end, and which will cause more
transactions to be included in flush.
was (Author: bharatviswa):
Hi [~rajesh.balamohan]
What is the test you are running? And also can we try with increasing handler
count to increase handler threads.
> Fix write performance issue in Non-HA OM
> -----------------------------------------
>
> Key: HDDS-2351
> URL: https://issues.apache.org/jira/browse/HDDS-2351
> Project: Hadoop Distributed Data Store
> Issue Type: Bug
> Components: om
> Reporter: Rajesh Balamohan
> Assignee: Nanda kumar
> Priority: Major
> Labels: performance
> Attachments: Screenshot 2019-10-23 at 2.27.05 PM.png
>
>
> HDDS-2333 enables sync option in OM non-HA mode. However, this flushes very
> frequently causing disk to saturate its IOPS soon. It creates way too small
> write workloads and disk hits limit.
> To put it in perspective, in simple write benchmark of creating keys with 10
> clients, it generates {{0.33MB/s}} write workload with {{116 IOPS}}. This
> causes disk to saturate at 98%.
>
> !Screenshot 2019-10-23 at 2.27.05 PM.png|width=621,height=370!
>
> Reverting back HDDS-2333 fixes this issue. I see >{{10x}} degradation with
> HDDS-2333.
> In case non-HA is supported in OM, it would be good to call it out.
> Currently, code explicitly enables sync option.
> [https://github.com/apache/hadoop-ozone/commit/c6c9794fc590371ad9c3b8fdcd7a36ed42909b40#diff-3ed3ab4891d7b4fa31ca96740b78ae5bR261]
>
> I used {{commit 1baa5a158d13f469c12bef86ef288d60ef0eee85}} in master branch.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]