On Thu, 3 Apr 2014, Tina Friedrich wrote:
...
Not what I meant - they can be attached to hosts, yes. They can't be
*consumed* on a per-host basis. Consumables are either per slot (when
set to 'yes') or per job (when set to 'job'). The latter works fine for
this when using array jobs (where each task uses one of the
consumables); however, if this were a large *multi-node* (as in MPI)
parallel job, it would count as one job, and thus use one consumable.
There's no type of resource to be used on a 'you will need to use X
amounts of these on every host you run on' type of consumable, if that
makes sense. (Had this problem before with GPUs...).
My track record of missing the point is quite high, so perhaps I ought to
butt out now. But I'll go on a bit more and show my ignorance once again
:)
Doesn't it depend on how you want your MPI job to be distributed?
For example, if it's a simple parallel code, you might only be interested
in your distribution of rank 0 processes on the cluster - so a per-job
consumable with a certain number configured on each host might do what you
want (but I've never used those - so I don't know from experience what
would happen there, especially if the parallel code is split across
nodes).
If it's a more sophisticated parallel code where each process contributes
to the loading you're worried about, you should be able to deal with it
using a per-slot consumable with a certain number configured on each host,
shouldn't you? That's the type I originally had in mind.
Or have I got the wrong end of the stick again?
Hope I explained it a bit better now... it took me a while to get my
head 'round this (actually, I think I had it explained to me by reuti on
this list a while ago...)
Reuti's great for that :)
Mark
--
-----------------------------------------------------------------
Mark Dixon Email : [email protected]
HPC/Grid Systems Support Tel (int): 35429
IT Services building Tel (ext): +44(0)113 343 5429
University of Leeds, LS2 9JT, UK
-----------------------------------------------------------------
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
