hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "He Tianyi (JIRA)" <j...@apache.org>
Subject [jira] [Comment Edited] (YARN-5510) Optimize Resource implementation for instantiation
Date Fri, 12 Aug 2016 05:45:20 GMT

    [ https://issues.apache.org/jira/browse/YARN-5510?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15418391#comment-15418391
] 

He Tianyi edited comment on YARN-5510 at 8/12/16 5:45 AM:
----------------------------------------------------------

A preliminary implementation. 
Tests is not included yet. Looking for suggestions & comments.

A simple benchmark in a simulated case with 2,000 nodes, 300 running apps:
*rewritten ResourceImplPB*
Total time elapsed: 19877 ms
Total number of heartbeats: 1773000
Cost per heartbeat: 11.210941906373378 us

*original ResourceImplPB*
Total time elapsed: 25909 ms
Total number of heartbeats: 1773000
Cost per heartbeat: 14.613085166384659 us



was (Author: he tianyi):
A preliminary implementation. 
Tests is not included yet. Looking for suggestions & comments.

> Optimize Resource implementation for instantiation
> --------------------------------------------------
>
>                 Key: YARN-5510
>                 URL: https://issues.apache.org/jira/browse/YARN-5510
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: resourcemanager
>    Affects Versions: 2.6.0
>            Reporter: He Tianyi
>            Assignee: He Tianyi
>         Attachments: YARN-5510.0001.patch
>
>
> According to YARN-5479, many {{Resource}} instances are created as intermediate result
of calculating in scheduler (as a result of {{Resources.subtract}} and {{Resources.add}}).
This introduces pressure on garbage collector as well as CPU cycles for scheduling.
> Therefore, the issue proposes a lightweight implementation of {{Resource}}, which optimized
for scheduler usage. That is, making instantiation of {{Resource}} cheaper.
> In discussion of {{YARN-5479}}, we thought maybe sticking to one version of implementation
avoids unnecessary complexity.
> In current implementation {{ResourceImplPB}}, a builder is also created during construction.
Also, {{maybeInitBuilder}} get called during each {{setMemory}} or {{setVirtualCores}}. The
{{builder}} get prepared regardless whether {{Resource}} will be exchanged via IPC or not.

> However, in scheduler (which performance is critical), most of these instances are just
intermediate result and not exposed.
> Suppose we could move builder related code into {{getProto}}.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org


Mime
View raw message