Ram Venkatesh commented on YARN-2362:

I agree that apps that need the entire cluster capacity  are likely not common. 
However, I think the scenario above can happen in busy clusters where an app 
might make a request that exceeds _current_ capacity and hence block all other 
apps. Yes, whenever more resources get freed up and App1's request is 
satisfied, only then will App2 run. Note, since we are enumerating the set of 
active apps, the behavior is actually non-deterministic - if the new app 
happens to be enumerated before the large app, the allocation request will 
actually be satisfied. The change proposed here makes it deterministic and can 
also reduce the wait for jobs that can complete - the downside of course is the 
large app can now experience starvation if small apps keep getting through. 

> Capacity Scheduler apps with requests that exceed capacity can starve pending 
> apps
> ----------------------------------------------------------------------------------
>                 Key: YARN-2362
>                 URL: https://issues.apache.org/jira/browse/YARN-2362
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: capacityscheduler
>    Affects Versions: 2.4.1
>            Reporter: Ram Venkatesh
> Cluster configuration:
> Total memory: 8GB
> yarn.scheduler.minimum-allocation-mb 256
> yarn.scheduler.capacity.maximum-am-resource-percent 1 (100%, test only config)
> App 1 makes a request for 4.6 GB, succeeds, app transitions to RUNNING state. 
> It subsequently makes a request for 4.6 GB, which cannot be granted and it 
> waits.
> App 2 makes a request for 1 GB - never receives it, so the app stays in the 
> ACCEPTED state for ever.
> I think this can happen in leaf queues that are near capacity.
> The fix is likely in LeafQueue.java assignContainers near line 861, where it 
> returns if the assignment would exceed queue capacity, instead of checking if 
> requests for other active applications can be met.
> {code:title=LeafQueue.java|borderStyle=solid}
>            // Check queue max-capacity limit
>            if (!assignToQueue(clusterResource, required)) {
> -            return NULL_ASSIGNMENT;
> +            break;
>            }
> {code}
> With this change, the scenario above allows App 2 to start and finish while 
> App 1 continues to wait.
> I have a patch available, but wondering if the current behavior is by design.

This message was sent by Atlassian JIRA

Reply via email to