Peng Zhang commented on YARN-3453:

I‘ve the same problem in our cluster. I think 2 points should be done to make 
it work:
# Calculator should be configurable, so isStarved() can use 
DominantResourceCalculator to test. But this still has corner case: 
cluster(40960, 8), expected( 16384, 4), real(4096, 4) it still result with 
"starved". My idea for this is to lower starve threshold to ignore this. Any 
good suggestions?
# in resToPreempt(), we should change logic of calculation, changing calculator 
is not enough. My hacked code is like below:
-      Resource target = Resources.min(getResourceCalculator(), clusterResource,
-          sched.getFairShare(), sched.getDemand());
+      Resource resourceUpperBound = Resources.componentwiseMin(
           sched.getFairShare(), sched.getDemand()); // min value of both cpu 
and memory as upperBound of request resource
+      float targetRatio = 0;
+        // getResourceAsValue is not super method in ResourceCalculator, and i 
cannot figure out one name for this calculation logic, so hack like this
+      if (resourceCalculator instanceof DominantResourceCalculator) {
+        // the ratio of demand can be get under resourceUpperBound.
+        targetRatio = ((DominantResourceCalculator)resourceCalculator)
+            .getResourceAsValue(sched.getDemand(), resourceUpperBound, false); 
 // min of ratio for cpu and memory
+      } else {
+        targetRatio = Resources.ratio(resourceCalculator, sched.getDemand(),
+            resourceUpperBound);
+      }
+      Resource target = Resources.multiply(sched.getDemand(), targetRatio);  
// demand resource can be fulfilled under fair share. 

Besides "drf" policy problem with preemption, I also filed YARN-3405 to 
describe some common problems with preemption.

> Fair Scheduler : Parts of preemption logic uses DefaultResourceCalculator 
> even in DRF mode causing thrashing
> ------------------------------------------------------------------------------------------------------------
>                 Key: YARN-3453
>                 URL: https://issues.apache.org/jira/browse/YARN-3453
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: fairscheduler
>    Affects Versions: 2.6.0
>            Reporter: Ashwin Shankar
> There are two places in preemption code flow where DefaultResourceCalculator 
> is used, even in DRF mode.
> Which basically results in more resources getting preempted than needed, and 
> those extra preempted containers aren’t even getting to the “starved” queue 
> since scheduling logic is based on DRF's Calculator.
> Following are the two places :
> 1. {code:title=FSLeafQueue.java|borderStyle=solid}
> private boolean isStarved(Resource share)
> {code}
> A queue shouldn’t be marked as “starved” if the dominant resource usage
> is >=  fair/minshare.
> 2. {code:title=FairScheduler.java|borderStyle=solid}
> protected Resource resToPreempt(FSLeafQueue sched, long curTime)
> {code}
> --------------------------------------------------------------
> One more thing that I believe needs to change in DRF mode is : during a 
> preemption round,if preempting a few containers results in satisfying needs 
> of a resource type, then we should exit that preemption round, since the 
> containers that we just preempted should bring the dominant resource usage to 
> min/fair share.

This message was sent by Atlassian JIRA

Reply via email to