[YARN-7560] Resourcemanager hangs when resourceUsedWithWeightToResourceRatio return a overflow value

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 3.0.0
• Component/s: fairscheduler, resourcemanager
• Target Version/s: 3.0.4
• Assignee: Chenyu Zheng
• Reporter: Chenyu Zheng
• Created: 26/Nov/17 03:45
• Updated: 26/Feb/20 05:29
• Resolved: 12/Jan/19 13:11

Description

In our cluster, we changed the configuration, then refreshQueues, we found the resourcemanager hangs. And the Resourcemanager can't restart successfully. We got jstack information, always show like this:

"main" #1 prio=5 os_prio=0 tid=0x00007f98e8017000 nid=0x2f5 runnable [0x00007f98eed9a000]
   java.lang.Thread.State: RUNNABLE
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.ComputeFairShares.resourceUsedWithWeightToResourceRatio(ComputeFairShares.java:182)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.ComputeFairShares.computeSharesInternal(ComputeFairShares.java:140)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.ComputeFairShares.computeSteadyShares(ComputeFairShares.java:66)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.policies.FairSharePolicy.computeSteadyShares(FairSharePolicy.java:148)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FSParentQueue.recomputeSteadyShares(FSParentQueue.java:102)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.QueueManager.getQueue(QueueManager.java:148)
        - locked <0x00007f8c4a8177a0> (a java.util.HashMap)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.QueueManager.getLeafQueue(QueueManager.java:101)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.QueueManager.updateAllocationConfiguration(QueueManager.java:387)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler$AllocationReloadListener.onReload(FairScheduler.java:1728)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.AllocationFileLoaderService.reloadAllocations(AllocationFileLoaderService.java:422)
        - locked <0x00007f8c4a7eb2e0> (a org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.AllocationFileLoaderService)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler.initScheduler(FairScheduler.java:1597)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler.serviceInit(FairScheduler.java:1621)
        at org.apache.hadoop.service.AbstractService.init(AbstractService.java:163)
        - locked <0x00007f8c4a76ac48> (a java.lang.Object)
        at org.apache.hadoop.service.CompositeService.serviceInit(CompositeService.java:107)
        at org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$RMActiveServices.serviceInit(ResourceManager.java:569)
        at org.apache.hadoop.service.AbstractService.init(AbstractService.java:163)
        - locked <0x00007f8c49254268> (a java.lang.Object)
        at org.apache.hadoop.yarn.server.resourcemanager.ResourceManager.createAndInitActiveServices(ResourceManager.java:997)
        at org.apache.hadoop.yarn.server.resourcemanager.ResourceManager.serviceInit(ResourceManager.java:257)
        at org.apache.hadoop.service.AbstractService.init(AbstractService.java:163)
        - locked <0x00007f8c467495e0> (a java.lang.Object)
        at org.apache.hadoop.yarn.server.resourcemanager.ResourceManager.main(ResourceManager.java:1220)

When we debug the cluster, we found resourceUsedWithWeightToResourceRatio return a negative value. So the loop can't return. We found in our cluster, the sum of all minRes is over int.max, so resourceUsedWithWeightToResourceRatio return a negative value.

below is the loop. Because totalResource is long, so always postive. But resourceUsedWithWeightToResourceRatio return int type. Our cluster is so big that resourceUsedWithWeightToResourceRatio will return a overflow value, just a negative. So the loop will never break.

    while (resourceUsedWithWeightToResourceRatio(rMax, schedulables, type)
        < totalResource) {
      rMax *= 2.0;
    }

Issue Links

• duplicates: YARN-9173 FairShare calculation broken for large values after YARN-8833