Am 20.12.2011 um 05:24 schrieb mahbube rustaee:

> Hi
> h_vmem set to physical memory on any node.
> I set h_vmem on 48 cpus hosts to 100G and submitted a job  to queue instance 
> on one host with requests:
> #$ -pe mpifill 48
> #$ -l  h_vmem=50M
> 
> $> qsub -q [email protected] detdet.job
> 
> while the job is running:
> $> qhost -F -h  amd-9-6.local 
> ...
> hc:h_vmem=97.656G
> 
> 1) all h_vmem request for that job is 48*50M=2400M   remaining h_vmem= 
> 97.656G 
> consumed h_vmem added to remaining h_vmem should equal be 100G but 
> 2400M+97.656G=100056 M (100.56G) that is greater than h_vmem on host level .
> job ran successfully . 
> How 56M could be described?

Hint: `man sge_types`.


> 2)  another job: 
> 
> #$ -pe mpi12amd 120     # PE config is 12 cores per node
> #$ -l  h_vmem=1.4G
> 
>   request value of h_vmem is strange!
>  I should set h_vmem for a job to 1.4G, otherwise run of it will be failed. 
> It means that job consumes
> 12* 1.4G memory on any node (16.8G). 
> 
> while job is running, <qhost -F -h host>  shows that  remaining mem_free 
> doesn't match with consuming h_vmem value.
> it means that job doesn't consumes such value as real memory  !!!
> 
> hl:mem_free=17.057G  #first value of mem_free is 23.7G , before of accept 
> this job
> hc:h_vmem=6.200G
> 
> consumed mem_free (23G-17.057G=6G) is less 10G than consumed h_vmem (16.8G).
>  why h_vmem should be requested more than It needed?

Depends on the application. And as said: the master process needs more memory.


> this value of h_vmem is different host to host! on xeon hosts (with 24cpus 
> and 23G h_vmem) requested h_vmem for that job  should be at least1G, and for 
> AMD hosts(24cpus and 100Gh_vmem) requested h_vmem for that job  should be at 
> least 1.4G!!! why?

SGE can't predict what will be used from the granted memory. Maybe the 
application uses to much memory during a malloc or whatever. As it's different 
from machine to machine: does always the same rank use the same amount and it's 
inside the application?

-- Reuti


> 
> Thx
> _______________________________________________
> users mailing list
> [email protected]
> https://gridengine.org/mailman/listinfo/users


_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to