Peng Zhang commented on YARN-3405:

And there is another related case which will cause live lock during preemption 
and scheduling.
If necessary, I will create a separated issue for it.

Queue hierarchy described as below:
              /        |        \
          queue-1    queue-2    queue-3     
          /    \
queue-1-1      queue-1-2

# Assume cluster resource is 100G in memory
# Assume queue-1 has max resource limit 20G
# queue-1-1 is active and it will get max 20G memory(equal to its fairshare)
# queue-2 is active then, and it require 30G memory(less than its fairshare)
# queue-3 is active, and it can be assigned with all other resources, 50G 
memory(larger than its fairshare)
# queue-1-2 is active, it will cause new preemption request(10G memory and 
intuitively it can only preempt from its sibling queue-1-1)
# Actually preemption starts from root, and it will find queue-3 is most over 
fairshare, and preempt some resources form queue-3.
# But during scheduling, it will find queue-1 itself arrived it's max 
fairshare, and cannot assign resource to it. Then resource's again assigned to 

And then it repeats between last two steps.

> FairScheduler's preemption cannot happen between sibling in some case
> ---------------------------------------------------------------------
>                 Key: YARN-3405
>                 URL: https://issues.apache.org/jira/browse/YARN-3405
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: fairscheduler
>    Affects Versions: 2.7.0
>            Reporter: Peng Zhang
>            Priority: Critical
> Queue hierarchy described as below:
> {noformat}
>              root
>               |         
>            queue-1        
>           /      \
> queue-1-1        queue-1-2
> {noformat}
> 1. When queue-1-1 is active and it has been assigned with all resources.
> 2. When queue-1-2 is active, and it cause some new preemption request.
> 3. But when do preemption, it now starts from root, and found queue-1 is not 
> over fairshare, so no recursion preemption to queue-1-1.
> 4. Finally queue-1-2 will be waiting for resource release form queue-1-1 
> itself.

This message was sent by Atlassian JIRA

Reply via email to