Hi, I suspect you have to use the appropriate scheduling algorithm.
Have a look at the SelectType parameter in 'man slurm.conf.

*/David*


On Tue, Mar 5, 2013 at 7:12 AM, Joo-Kyung Kim <[email protected]> wrote:

> Hi,****
>
> ** **
>
> I am using SLURM 2.4.0, and I have three nodes, each of which has a GPU.**
> **
>
> ** **
>
> Currently, a job requested a gpu by “srun –n 1 –-gres=gpu --multi-prog
> <script1>” is running. While that job is running, I requested another job
> by “srun –n 40 --gres=gpu --multi-prog <script2>”.****
>
> ** **
>
> In this case, because two gpus are available, I expected the second job
> would run using the two remaining gpus. However, the second job is pended
> as follows.****
>
> JOBID PARTITION     NAME     USER  ST       TIME  NODES NODELIST(REASON)**
> **
>
> 26646 xxxxxxxxx script1  kimjook  PD       0:00      2 (Resources)****
>
> 26645 xxxxxxxxx script2  kimjook   R      14:23      1 node1****
>
> ** **
>
> When the first job is finished, then the second job could run using all
> three gpus.****
>
> ** **
>
> So, if there is a running job occupied a gpu, and another job is
> requesting gpus more than the number of gpus in the cluster, I wonder if
> there is a way that the second job can run using the two remaining gpus.**
> **
>
> ** **
>
> Thanks.****
>
> ** **
>
> Joo-Kyung****
>

Reply via email to