[gridengine users] problem with PE / complex setup - not sure, this might even be a bug?
d.love at liverpool.ac.uk
Tue Jul 16 22:23:51 UTC 2013
Tina Friedrich <Tina.Friedrich at diamond.ac.uk> writes:
> qstat -F gpus=*,slots=* explains why, I think. The nodes have 2 GPUs
> available (qstat -F confirms that). Now, when I submit asking for 1
> GPU and 2 slots (i.e. '-l gpus=1 -pe smp 2'), I'd expect the there to
> be 1 GPU resource and 2 slots used. However, qstat shows 2 GPU
> resources used (i.e. non left) - and two slots used, as expected.
> I just tried with another consumable (called matlab) I've introduced;
> same sort of thing happens - the nodes have 1 available; scheduling as
> -l matlab -pe smp 8' ends me up with the node showing "-7" matlab
> resources left (I'd expect 0 left).
> So somehow, it seems that the number of slots asked for (via the PE)
> gets also uses as the number of GPUs requested?
> Is that me doing something stupid in my definitions, or a bug? Pretty
> sure this worked "as expected" in SGE6.2u5.
It was the same in 6.2u5 (see complex(5)). I guess you really want
per-host consumables, which aren't currently implemented. If only
single node jobs are relevant, look at JOB consumables.
I was trying to find out recently how the old Sun remote visualization
stuff worked and whether it did anything particularly clever for
handling the GPUs. Can anyone say (not that I expect it was really good
Community Grid Engine: http://arc.liv.ac.uk/SGE/
More information about the users