Slurm is designed to dedicate resources (e.g. CPUs, memory, GPUs,  
etc.) and bind those resources to specific user jobs. Slurm does  
record system load and this could be used as a basis of scheduling  
decisions with minor code changes, but that will probably severely  
impact parallel job performance, which is what Slurm was really  
designed for.

Quoting Mario Kadastik <[email protected]>:

>
> Hi,
>
> is it possible to configure SLURM to use system load as a basis on  
> where to schedule as well as when to stop sending new jobs to shared  
> resources?
>
> The reason I ask is that we have 24 core and 32 core nodes and  
> 99.99% of our jobs are single core jobs. So we'd configure SLURM  
> with consumable resources and allow therefore 24 or 32 jobs to share  
> the nodes. We'd however like it to be done so that if say we had 170  
> work servers the first 170 jobs would each get one of the servers,  
> the next 170 would become the second job in each server etc.
>
> The second part is that we'd like during transition period run slurm  
> and torque in parallel. This means that we've already set up torque  
> to schedule based on system load and pbs_mom has an attribute of  
> max_load that is adhered to therefore if the system load rises above  
> 10% of the core count (27 for 24 and 36 for 32) there are no more  
> jobs scheduled as it's assumed the node is busy. It can lead to even  
> blocking the scheduling of further jobs even though count wise there  
> may be free slots still. Can the same be done for slurm? This way  
> the two could schedule jobs both and only if the cluster really gets  
> full would both stop scheduling until some jobs end and the load  
> drops below some margin.
>
> If that can be achieved, then can you let me know what I'd need in  
> the slurm.conf to make it work like this :) Or is that the default  
> behavior.
>
> Thanks,
>
> Mario Kadastik, PhD
> Researcher
>
> ---
>   "Physics is like sex, sure it may have practical reasons, but  
> that's not why we do it"
>      -- Richard P. Feynman
>

Reply via email to