[
https://issues.apache.org/jira/browse/YARN-5982?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15733021#comment-15733021
]
Arun Suresh commented on YARN-5982:
-----------------------------------
Thanks for the patch [~kkaranasos]
Couple of comments:
# You can probably completely remove the
OpportunisticContainerAllocatorAMService::createMin/MaxResource, and replace it
with {{rmContext.getScheduler().getMinimumResourceCapability()}} and
{{rmContext.getScheduler().getMaximumResourceCapability()}}. This will ensure
the values are in sync with the actual scheduler being used.
# Lets not remove the incr resources from the
{{yarn_server_common_service_protos.proto}}, since there exists a
*yarn.scheduler.increment-allocation-mb/vcore*. It is used by the
FairScheduler, but not the capacity scheduler yet, and the normalize function
actually has an argument for it. I feel, it would be better if we just add
*getIncrementResourceCapability()* in YarnScheduler interface, and in
AbstractYarnScheduler, put in a default implementation that delegates to
*getMinimumResourceCapability()*. FairScheduler already has this, so you just
have to put an *@Override* tag on the method.
# In {{NodeManagerMetrics}}, if you are changing the
*opportunisticContainerStarted* method to *startOpportunisticContainer*, you
should also correspondingly change *opportunisticContainerCompleted* to
*completeOpportunisticContainer*. Also, lets no not move the placement of the
methods, since it can cause merge issues during cherry-pick and the diffs
become larger and difficult to make sense of sometimes.
> Simplifying opportunistic container parameters and metrics
> ----------------------------------------------------------
>
> Key: YARN-5982
> URL: https://issues.apache.org/jira/browse/YARN-5982
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Konstantinos Karanasos
> Assignee: Konstantinos Karanasos
> Fix For: 2.9.0, 3.0.0-alpha2
>
> Attachments: YARN-5982.001.patch
>
>
> This JIRA removes some of the parameters that are related to opportunistic
> containers (e.g., min/max memory/cpu). Instead, we will be using the
> parameters already used by guaranteed containers.
> The goal is to reduce the number of parameters that need to be used by the
> user.
> We also fix a small issue related to the container metrics (opportunistic
> memory reported in GB in Web UI, although it was in MB).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
