Oxana Smirnova <[log in to unmask]> writes:
> Leif Nixon пишет:
>>
>> Those memory limits are a mixed blessing; our nodes don't run out of
>> memory, but it's a bit of a pity that jobs are being killed when they
>> exceed their asked-for memory limit of, say, 800 MB, when they happen
>> to be running on a dedicated node with 2 GB RAM.
>>
>
> Leif, it *is* a blessing: if such job keeps running (I have to admit
> that I increased this limit to 1300 MB two days ago), it gets stuck and
> occupies the processor not for 30 minutes, but for 5 hours, doing
> nothing (well, filling up the disk with garbage messages perhaps), only
> to be killed by the requested (and correctly enforced) CPU time
> consumption limit.
Oh, yeah, I wouldn't like to be without it. But perhaps the enforced
limit should be set to a fixed value; the total amount of memory in a
node minus a bit. (If you have job-dedicated nodes, that is.) Use the
memory specified in the job description for brokering purposes only.
--
Leif Nixon - Systems expert
------------------------------------------------------------
National Supercomputer Centre - Linkoping University
------------------------------------------------------------
|