[ 
https://issues.apache.org/jira/browse/YARN-10293?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17124195#comment-17124195
 ] 

Wangda Tan commented on YARN-10293:
-----------------------------------

[~Tao Yang], the suggestion totally make sense to me. When we have done the 
initial global scheduling framework, the goal is to make it compatible to the 
previous behavior, I agree to make additional steps to overhaul reservation 
logic under the context of global scheduling is a good idea. Now the code is 
very hard to read and understand.

I think we can do this step by step, first, let's fix low hanging fruits like 
this Jira. (I hope to get idea from you about the proposed change: 
https://issues.apache.org/jira/browse/YARN-10293?focusedCommentId=17121419&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-17121419

And [~prabhujoseph] if you have time/bandwidth, can you take a look into 
reservation related logic + preemption + unreserve + global scheduling and see 
what we can optimize here?

 

> Reserved Containers not allocated from available space of other nodes in 
> CandidateNodeSet in MultiNodePlacement (YARN-10259)
> ----------------------------------------------------------------------------------------------------------------------------
>
>                 Key: YARN-10293
>                 URL: https://issues.apache.org/jira/browse/YARN-10293
>             Project: Hadoop YARN
>          Issue Type: Bug
>    Affects Versions: 3.3.0
>            Reporter: Prabhu Joseph
>            Assignee: Prabhu Joseph
>            Priority: Major
>         Attachments: YARN-10293-001.patch, YARN-10293-002.patch, 
> YARN-10293-003-WIP.patch
>
>
> Reserved Containers not allocated from available space of other nodes in 
> CandidateNodeSet in MultiNodePlacement. YARN-10259 has fixed two issues 
> related to it 
> https://issues.apache.org/jira/browse/YARN-10259?focusedCommentId=17105987&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17105987
> Have found one more bug in the CapacityScheduler.java code which causes the 
> same issue with slight difference in the repro.
> *Repro:*
> *Nodes :       Available :         Used*
> Node1 -  8GB, 8vcores -  8GB. 8cores
> Node2 -  8GB, 8vcores - 8GB. 8cores
> Node3 -  8GB, 8vcores - 8GB. 8cores
> Queues -> A and B both 50% capacity, 100% max capacity
> MultiNode enabled + Preemption enabled
> 1. JobA submitted to A queue and which used full cluster 24GB and 24 vcores
> 2. JobB Submitted to B queue with AM size of 1GB
> {code}
> 2020-05-21 12:12:27,313 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.RMAuditLogger: USER=systest  
> IP=172.27.160.139       OPERATION=Submit Application Request    
> TARGET=ClientRMService  RESULT=SUCCESS  APPID=application_1590046667304_0005  
>   CALLERCONTEXT=CLI       QUEUENAME=dummy
> {code}
> 3. Preemption happens and used capacity is lesser than 1.0f
> {code}
> 2020-05-21 12:12:48,222 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptMetrics:
>  Non-AM container preempted, current 
> appAttemptId=appattempt_1590046667304_0004_000001, 
> containerId=container_e09_1590046667304_0004_01_000024, 
> resource=<memory:1024, vCores:1>
> {code}
> 4. JobB gets a Reserved Container as part of 
> CapacityScheduler#allocateOrReserveNewContainer
> {code}
> 2020-05-21 12:12:48,226 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainerImpl: 
> container_e09_1590046667304_0005_01_000001 Container Transitioned from NEW to 
> RESERVED
> 2020-05-21 12:12:48,226 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp:
>  Reserved container=container_e09_1590046667304_0005_01_000001, on node=host: 
> tajmera-fullnodes-3.tajmera-fullnodes.root.hwx.site:8041 #containers=8 
> available=<memory:0, vCores:0> used=<memory:8192, vCores:8> with 
> resource=<memory:1024, vCores:1>
> {code}
> *Why RegularContainerAllocator reserved the container when the used capacity 
> is <= 1.0f ?*
> {code}
> The reason is even though the container is preempted - nodemanager has to 
> stop the container and heartbeat and update the available and unallocated 
> resources to ResourceManager.
> {code}
> 5. Now, no new allocation happens and reserved container stays at reserved.
> After reservation the used capacity becomes 1.0f, below will be in a loop and 
> no new allocate or reserve happens. The reserved container cannot be 
> allocated as reserved node does not have space. node2 has space for 1GB, 
> 1vcore but CapacityScheduler#allocateOrReserveNewContainers not getting 
> called causing the Hang.
> *[INFINITE LOOP] CapacityScheduler#allocateContainersOnMultiNodes -> 
> CapacityScheduler#allocateFromReservedContainer -> Re-reserve the container 
> on node*
> {code}
> 2020-05-21 12:13:33,242 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler:
>  Trying to fulfill reservation for application application_1590046667304_0005 
> on node: tajmera-fullnodes-3.tajmera-fullnodes.root.hwx.site:8041
> 2020-05-21 12:13:33,242 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.LeafQueue: 
> assignContainers: partition= #applications=1
> 2020-05-21 12:13:33,242 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp:
>  Reserved container=container_e09_1590046667304_0005_01_000001, on node=host: 
> tajmera-fullnodes-3.tajmera-fullnodes.root.hwx.site:8041 #containers=8 
> available=<memory:0, vCores:0> used=<memory:8192, vCores:8> with 
> resource=<memory:1024, vCores:1>
> 2020-05-21 12:13:33,243 INFO 
> org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler:
>  Allocation proposal accepted
> {code}
> CapacityScheduler#allocateOrReserveNewContainers won't be called as below 
> check in allocateContainersOnMultiNodes fails
> {code}
>  if (getRootQueue().getQueueCapacities().getUsedCapacity(
>         candidates.getPartition()) >= 1.0f
>         && preemptionManager.getKillableResource(
> {code}



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org

Reply via email to