The current SLURM GRES logic was written to allocate the same number of those generic resources per node rather than a total count of resources across an entire job or step allocation. If you want to allocate some total count of GRES, I would suggest looking at the current SLURM code and data structures with the thought of modifying them to add this functionality. The code is not simple, but that would probably be the easiest way to proceed. Most of the code is in src/common/gres.c.
Quoting Sergio Iserte <[email protected]>: > Well, I need to implement a plug-in for this behaviour, but I don't know > how to start. > I was thinking of a plug-in which can allocate individual GPUs as does the > select/cons_res. > And then, to implement a scheduler plug-in for GPUs. > How about this idea? > > Regars! > > On Monday, 21 March 2011 05:37:53 UTC+1, Moe Jette wrote: >> >> You can specify 2 GPUs per node scheduled, but not a mixed count >> today and no change is planned right now to do that. >> >> Quoting Ahmet Baltac? <[email protected]>: >> >> > Dear all, >> > >> > I'm currently investigating SLURM's features, and there is a point >> > that I couldn't be sure of. >> > >> > Lets say, in a partition, node A has 3 GPU's defined as gres, and node >> > B has 2 GPU's defined as gres. Is there any way to schedule a job with >> > a requirement of 5 GPU's ? >> > >> > Thanks, >> > Ahmet >> > >> >>
