[
https://issues.apache.org/jira/browse/YARN-1458?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14107066#comment-14107066
]
zhihai xu commented on YARN-1458:
---------------------------------
[~shurong.mai], YARN-1458.patch will cause regression. It won't work if all the
weight and MinShare in the active queues are less than 1.
The type conversion from double to int in computeShare loses precision.
{code}
private static int computeShare(Schedulable sched, double w2rRatio,
ResourceType type) {
double share = sched.getWeights().getWeight(type) * w2rRatio;
share = Math.max(share, getResourceValue(sched.getMinShare(), type));
share = Math.min(share, getResourceValue(sched.getMaxShare(), type));
return (int) share;
}
{code}
In above code, the initial value w2rRatio is 1.0. If weight and MinShare are
less than 1, computeShare will return 0.
resourceUsedWithWeightToResourceRatio will return the sum of all these return
values from computeShare(after lose precision).
It will be zero if all the weight and MinShare in the active queues are less
than 1. Then YARN-1458.patch will exit the loop earlier with
"rMax" value 1.0. Then "right" variable will be less than "rMax"(1.0). Then all
queues' fair share will be set to 0 in the following code.
{code}
for (Schedulable sched : schedulables) {
setResourceValue(computeShare(sched, right, type), sched.getFairShare(),
type);
}
{code}
This is the reason why the TestFairScheduler is failed at line 1049.
testIsStarvedForFairShare configure the queueA weight 0.25 and queueB weight
0.75 and total node resource 4 * 1024.
It creates two applications: one is assigned to queueA and the other is
assigned to queueB.
After FaiScheduler(update) calculated the fair share, queueA fair share should
be 1 * 1024 and queueB fair share should be 3 * 1024.
but with YARN-1458.patch, both queueA fair share and queueB fair share are set
to 0,
It is because in this test there are two active queues:queueA and queueB, both
weights are less than 1(0.25 and 0.75), MinShare(minResources) in queueA and
queueB are not configured, both MinShare use default value(0).
> In Fair Scheduler, size based weight can cause update thread to hold lock
> indefinitely
> --------------------------------------------------------------------------------------
>
> Key: YARN-1458
> URL: https://issues.apache.org/jira/browse/YARN-1458
> Project: Hadoop YARN
> Issue Type: Bug
> Components: scheduler
> Affects Versions: 2.2.0
> Environment: Centos 2.6.18-238.19.1.el5 X86_64
> hadoop2.2.0
> Reporter: qingwu.fu
> Assignee: zhihai xu
> Labels: patch
> Fix For: 2.2.1
>
> Attachments: YARN-1458.001.patch, YARN-1458.002.patch, YARN-1458.patch
>
> Original Estimate: 408h
> Remaining Estimate: 408h
>
> The ResourceManager$SchedulerEventDispatcher$EventProcessor blocked when
> clients submit lots jobs, it is not easy to reapear. We run the test cluster
> for days to reapear it. The output of jstack command on resourcemanager pid:
> {code}
> "ResourceManager Event Processor" prio=10 tid=0x00002aaab0c5f000 nid=0x5dd3
> waiting for monitor entry [0x0000000043aa9000]
> java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler.removeApplication(FairScheduler.java:671)
> - waiting to lock <0x000000070026b6e0> (a
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler.handle(FairScheduler.java:1023)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler.handle(FairScheduler.java:112)
> at
> org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$SchedulerEventDispatcher$EventProcessor.run(ResourceManager.java:440)
> at java.lang.Thread.run(Thread.java:744)
> ……
> "FairSchedulerUpdateThread" daemon prio=10 tid=0x00002aaab0a2c800 nid=0x5dc8
> runnable [0x00000000433a2000]
> java.lang.Thread.State: RUNNABLE
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler.getAppWeight(FairScheduler.java:545)
> - locked <0x000000070026b6e0> (a
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.AppSchedulable.getWeights(AppSchedulable.java:129)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.ComputeFairShares.computeShare(ComputeFairShares.java:143)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.ComputeFairShares.resourceUsedWithWeightToResourceRatio(ComputeFairShares.java:131)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.ComputeFairShares.computeShares(ComputeFairShares.java:102)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.FairSharePolicy.computeShares(FairSharePolicy.java:119)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FSLeafQueue.recomputeShares(FSLeafQueue.java:100)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FSParentQueue.recomputeShares(FSParentQueue.java:62)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler.update(FairScheduler.java:282)
> - locked <0x000000070026b6e0> (a
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler)
> at
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler$UpdateThread.run(FairScheduler.java:255)
> at java.lang.Thread.run(Thread.java:744)
> {code}
--
This message was sent by Atlassian JIRA
(v6.2#6252)
