[Mauiusers] node allocation issue

Jeff Anderson-Lee jonah at eecs.berkeley.edu
Thu Dec 2 10:53:54 MST 2010


On 12/2/2010 9:35 AM, Abhishek Gupta wrote:
> Hi Jeff,
> Yes I did verified that multiple jobs are running on same logical cpu.
> Thanks,
> Abhi.

What is your NODEAVAILABILITYPOLICY set to?   We use:

NODEAVAILABILITYPOLICY COMBINED:MEM COMBINED:PROC COMBINED:SWAP

Also check out NODELOADPOLICY and JOBNODEMATCHPOLICY.

Jeff

> Jeff Anderson-Lee wrote:
>> On 12/2/2010 9:07 AM, Abhishek Gupta wrote:
>>> Our job scheduler is acting little weird. The jobs submitted are going
>>> on only a few nodes and each node is running more jobs than the total
>>> number of CPUs on that node. We have not changed anything recently.
>>> Can anyone suggest what is the best node allocation policy that make
>>> sure we do not run more jobs than number of cores available on a node
>>> and jobs are evenly distributed on all the nodes and not on the few 
>>> nodes?
>>> Thanks,
>>> Abhi.
>>
>> We find that many users understate the number of cores/procs needed 
>> in their requests.  They ask for a single cpu, then spawn multiple 
>> tasks, or run Java in a way that runs two or more active threads 
>> instead of one.  Maui does not seem to set any cpu affinity to 
>> prevent interference nor enforce the requested number of cores as far 
>> as I know.
>>
>> Do you know for certain that you are getting multiple jobs allocated 
>> to the same logical cpu and not just user task forks?
>>
>> If you use Torque as your underlying scheduler, the following should 
>> help you to find true duplicate assignments:
>>
>> qstat -f | grep exec | sort | uniq -c | sort -n
>>
>> Jeff Anderson-Lee



More information about the mauiusers mailing list