[ 
https://issues.apache.org/jira/browse/YARN-9015?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16707587#comment-16707587
 ] 

Wangda Tan commented on YARN-9015:
----------------------------------

[~tangzhankun], 

1) DevicePluginScheduler: Why use Integer instead int?

2) DeviceMappingManager:
- devicePluginSchedulers: suggest to make it accessed under synchronized, given 
assignDevices is accessed under synchronized lock. DeviceMappingManager is not 
used under highly concurrency env. Suggest to make all access to synchronized 
to simplify logics. 
- 
{code}
312           // TODO: should check if customized scheduler return values are 
valid
313           if (dpsAllocated.size() != count) {
{code} 
Is there any check needed?
- 
{code} 
        // TODO: fall back to default schedule logic?
{code} 
I think your existing throw exception logic is good enough. 


> Phase 1 - Add an interface for device plugin to provide customized scheduler
> ----------------------------------------------------------------------------
>
>                 Key: YARN-9015
>                 URL: https://issues.apache.org/jira/browse/YARN-9015
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>            Reporter: Zhankun Tang
>            Assignee: Zhankun Tang
>            Priority: Major
>         Attachments: YARN-9015-trunk.001.patch, YARN-9015-trunk.002.patch, 
> YARN-9015-trunk.003.patch
>
>
> A vendor might need a customized scheduling policy for their devices. It 
> could be scheduled based on topology, resource utilization, virtualization, 
> device attribute and so on.
> We'll provide another optional interface "DevicePluginScheduler" for the 
> vendor device plugin to implement. Once it's implemented, the framework will 
> prefer it to the default scheduler.
> This would bring more flexibility to the framework's scheduling mechanism.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

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

Reply via email to