Fair scheduler does not share resources of other queues

[Ilya Karpov <id...@gmail.com>]

Hi, all,

recently I’ve noticed strange behaviour of YARN Fair Scheduler: 2 jobs (i.e. two simultaneously started oozie launchers) started in a queue with a small weight, and was not able to launch spark jobs while there were plenty resources in other queues.

In details:
- hadoop(2.6, cdh 5.12) yarn with fair scheduler
- the queue(say, small_queue) with small weight starts 2 oozie launcher jobs
- oozie launcher jobs occupies all small_queue capacity (even exceed it by 1 core), and both ready to submit spark job (in the same queue = small_queue)
- there is about 1/4 of free cluster resources in other queues (much more then spark jobs require)

Expected behaviour: free resources of other queues will be given too oozie launchers (from small_queue) to start their spark jobs
Actual behaviour: spark jobs were never started

Does anybody have an idea what prevented spark jobs from launch?

Re: Fair scheduler does not share resources of other queues

[Gurmukh Singh <gu...@yahoo.com.INVALID>]

How is the pre-emption configured ?

On 21/4/20 1:41 am, Ilya Karpov wrote:
> Hi, all,
>
> recently I’ve noticed strange behaviour of YARN Fair Scheduler: 2 jobs 
> (i.e. two simultaneously started oozie launchers) started in a queue 
> with a small weight, and was not able to launch spark jobs while there 
> were plenty resources in other queues.
>
> In details:
> - hadoop(2.6, cdh 5.12) yarn with fair scheduler
> - the queue(say, *small_queue*) with small weight starts 2 oozie 
> launcher jobs
> - oozie launcher jobs occupies all small_queue capacity (even exceed 
> it by 1 core), and both ready to submit spark job (in the same queue = 
> small_queue)
> - there is about 1/4 of free cluster resources in other queues (much 
> more then spark jobs require)
>
> Expected behaviour: free resources of other queues will be given too 
> oozie launchers (from small_queue) to start their spark jobs
> Actual behaviour: spark jobs were never started
>
> Does anybody have an idea what prevented spark jobs from launch?