[YARN-8566] Add diagnostic message for unschedulable containers

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.2.0
• Component/s: resourcemanager
• Hadoop Flags: Reviewed
• Assignee: Szilard Nemeth
• Reporter: Szilard Nemeth
• Created: 23/Jul/18 13:21
• Updated: 27/Jul/18 21:36
• Resolved: 27/Jul/18 21:36

Description

If a queue is configured with maxResources set to 0 for a resource, and an application is submitted to that queue that requests that resource, that application will remain pending until it is removed or moved to a different queue. This behavior can be realized without extended resources, but it’s unlikely a user will create a queue that allows 0 memory or CPU. As the number of resources in the system increases, this scenario will become more common, and it will become harder to recognize these cases. Therefore, the scheduler should indicate in the diagnostic string for an application if it was not scheduled because of a 0 maxResources setting.

Example configuration (fair-scheduler.xml) :

<allocations>
  <queueMaxAppsDefault>100000</queueMaxAppsDefault>
<queue name="sample_queue">
    <minResources>10000 mb,2vcores</minResources>
    <maxResources>90000 mb,4vcores, 0gpu</maxResources>
    <maxRunningApps>50</maxRunningApps>
    <maxAMShare>-1.0f</maxAMShare>
    <weight>2.0</weight>
    <schedulingPolicy>fair</schedulingPolicy>
  </queue>
</allocations>

Command:

yarn jar "./share/hadoop/mapreduce/hadoop-mapreduce-examples-3.2.0-SNAPSHOT.jar" pi -Dmapreduce.job.queuename=sample_queue -Dmapreduce.map.resource.gpu=1 1 1000;

The job hangs and the application diagnostic info is empty.
Given that an exception is thrown before any mapper/reducer container is created, the diagnostic message of the AM should be updated.