Am 16.12.2011 um 09:05 schrieb mahbube rustaee: > Greetings, > > Yes, I referenced to the man page but : > > request h_vmem is per slot whereas h_vmem for host is host level and > for all slots. > In this situation, h_vmem job's requested should compare with h_vmem > host divide by all slots. > > the following relation must be checked for any job: > requested h_vmem <= host level h_vmem / $num_proc
Did you made the complex consumable (man complex)? Then it should work automatically this way. -- Reuti > Can I do that via RQS? > any suggestion would be appriciated. > > > Regards > Rustaee > > On 12/14/11, Reuti <[email protected]> wrote: >> Am 14.12.2011 um 14:00 schrieb mahbube rustaee: >> >>> for a heterogeneous cluster, that is different . >>> e.g. for AMD is 1G and for xeon is 700M. >>> If job will be scheduled on amd, be prevented to request more 1G on >>> h_vmem and for xeon >>> job be prevented to request 700M. >> >> There is a man page for defining the RQS: man sge_resource_quota >> >> -- Reuti >> >> >>> How can I do this? >>> Thx >>> >>> On Wed, Dec 14, 2011 at 4:08 PM, Reuti <[email protected]> wrote: >>> Am 14.12.2011 um 08:35 schrieb mahbube rustaee: >>> >>>> Hi Mr. Reuti , >>>> I set h_vmem equal to 48G and nodes have 48 slots (1G per slot). >>>> I mean that users request h_vmem at most 1G(this value is h_vmem per >>>> slot).If users can request more than 1G, some cores will be locked. I >>>> want to prevent such users at submission time to not happen locking >>>> other cores. >>> >>> So you want to limit the maximum one can request to be 1 GB? >>> >>> This can be achieved by a JSV or an RQS: >>> >>> limit to h_vmem=1G >>> >>> To avoid a successful submission with an endless waiting job, "-w e" as >>> default request can be used. >>> >>> -- Reuti >>> >>> >>>> Thx >>>> >>>> >>>> >>>> On Tue, Dec 13, 2011 at 9:57 PM, Reuti <[email protected]> >>>> wrote: >>>> Hi, >>>> >>>> Am 13.12.2011 um 11:46 schrieb mahbube rustaee: >>>> >>>>> Hi all, >>>>> I 'm so thankful in advance. >>>>> I set h_vmem to physical memory on host level. >>>>> suppose current state of a node with 4G h_vmem and 4 free slots. >>>>> a job with -l h_vmem=2G and pe mpi 2 launch to scheduler. >>>>> In this case job is running , h_vmem on that node will be zero and 2 >>>>> cores of the node can't be used! >>>> >>>> Yep, this is the purpose of making it consumable. What behavior would >>>> you like - when the job needs 4 GB, other jobs can't run there. Why >>>> would you like to run additional jobs there? >>>> >>>> >>>>> 2) How can I limit h_vmem request for any node at submission time ( >>>>> whereas h_vmem on host is total and request is per slots for any >>>>> job)? >>>> >>>> It's not implemented to have it one time per host. It was foreseen when >>>> the complexes got the new attribute "JOB", but it was never implemented >>>> for now. You will have to divide it by hand beforehand. >>>> >>>> I also don't understand you request, you want to submit with "-l >>>> h_vmem=2G", ok. It's per slot. So it will be multiplied and the result >>>> subtracted from the available memory. So it's per host already. >>>> >>>> -- Reuti >>>> >>> >>> >> >> _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
