Jason Lowe commented on YARN-3136:

Sorry for the delay in replying.  It is true that we could break some 
schedulers by moving the definition from a Map to a ConcurrentMap.  
Alternatives I see to that approach are:

1) use a read-write lock around accessing the map, which unfortunately can be 
wasteful for some map implementations (e.g.: concurrent maps).

2) use methods to access the map whose default implementation use a lock by 
default (for backwards compatibility) but can be overridden in derived 
schedulers for performance when the underlying map does not require a lock

3) use RTTI when accessing the map to see if it's concurrent and lock when it's 
not.  Not a huge fan of this one.

> getTransferredContainers can be a bottleneck during AM registration
> -------------------------------------------------------------------
>                 Key: YARN-3136
>                 URL: https://issues.apache.org/jira/browse/YARN-3136
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: scheduler
>    Affects Versions: 2.6.0
>            Reporter: Jason Lowe
>            Assignee: Sunil G
>         Attachments: 0001-YARN-3136.patch, 0002-YARN-3136.patch
> While examining RM stack traces on a busy cluster I noticed a pattern of AMs 
> stuck waiting for the scheduler lock trying to call getTransferredContainers. 
>  The scheduler lock is highly contended, especially on a large cluster with 
> many nodes heartbeating, and it would be nice if we could find a way to 
> eliminate the need to grab this lock during this call.  We've already done 
> similar work during AM allocate calls to make sure they don't needlessly grab 
> the scheduler lock, and it would be good to do so here as well, if possible.

This message was sent by Atlassian JIRA

Reply via email to