You are viewing a plain text version of this content. The canonical link for it is here.
Posted to yarn-issues@hadoop.apache.org by "Ahmed Hussein (JIRA)" <ji...@apache.org> on 2019/06/03 23:42:00 UTC
[jira] [Commented] (YARN-9597) Memory efficiency in speculator
[ https://issues.apache.org/jira/browse/YARN-9597?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16855155#comment-16855155 ]
Ahmed Hussein commented on YARN-9597:
-------------------------------------
By looking at previous Jira reports, it is clear that "memory leakage" is a persistent bug.
My concern is how many Hadoop modules have hidden bloating data structures? If this is a common issue, would it be useful to investigate test-cases/tools to detect memory leaks in Hadoop?
[~stevel@apache.org] are you aware of ongoing efforts in that direction?
> Memory efficiency in speculator
> --------------------------------
>
> Key: YARN-9597
> URL: https://issues.apache.org/jira/browse/YARN-9597
> Project: Hadoop YARN
> Issue Type: Improvement
> Reporter: Ahmed Hussein
> Priority: Minor
>
> The data structures in speculator and runtime-estimator are bloating. Data elements such as (taskID, TA-ID, task stats, tasks speculated, tasks finished..etc) are added to the concurrent maps but never removed.
> For long running jobs, there are couple of issues:
> # memory leakage: the speculator memory usage increases over time.
> # performance: keeping large structures in the heap affects the performance due to locality and cache misses.
> *Suggested Fixes:*
> - When a TA transitions to {{MoveContainerToSucceededFinishingTransition}}, the TA notifies the speculator. The latter handles the event by cleaning the internal structure accordingly.
> - When a task transitions is failed/killed, the speculator is notified to clean the internal data structure.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org