Tao Yang created YARN-9043:
------------------------------
Summary: Inter-queue preemption sometimes starves an underserved
queue when using DominantResourceCalculator
Key: YARN-9043
URL: https://issues.apache.org/jira/browse/YARN-9043
Project: Hadoop YARN
Issue Type: Bug
Components: capacityscheduler
Affects Versions: 3.3.0
Reporter: Tao Yang
Assignee: Tao Yang
To reproduce this problem in UT, we can setup a cluster with resource <40,18>
and create 3 queues and apps:
* queue a: guaranteed=<10,10>, used=<6,10> by app1
* queue b: guaranteed=<20,6>, used=<20,8> by app2
* queue c: guaranteed=<10,2>, used=<0,0>, pending=<1,1>
Queue c is an underserved queue, queue b overuses 2 cpu resource, we expect
app2 in queue b can be preempted but nothing happens.
This problem is related to Resources#greaterThan/lessThan, comparation between
two resources is based on the resource/cluster-resource ratio inside
DominantResourceCalculator#compare, in this way, the low weight resource may be
ignored, for the scenario in UT, take comparation between ideal assgined
resource and used resource:
* cluster resource is <40,18>
* ideal assigned resource of queue b is <20,6>, ideal-assigned-resource /
cluster-resource = <20, 6> / <40, 18> = max(20/40, 6/18) = 0.5
* used resource of queue b is <20, 8>, used-resource / cluster-resource = <20,
8> / <40, 18> = max(20/40, 8/18) = 0.5
The results of {{Resources.greaterThan(rc, clusterResource, used,
idealAssigned)}} will be false instead of true, and there are some other
similar places have the same problem, so that preemption can't happen in
current logic.
To solve this problem, I propose to add
ResourceCalculator#isAnyMajorResourceGreaterThan method, inside
DominantResourceCalculator implements, it will compare every resource type
between two resources and return true if any major resource types of left
resource is greater than that of right resource, then replace
Resources#greaterThan with it in some places of inter-queue preemption with
this problem.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]