[torquedev] nodes, procs, tpn and ncpus

[Glen Beane]

On Jun 10, 2010, at 3:09 PM, Garrick Staples <garrick at usc.edu> wrote:

> On Thu, Jun 10, 2010 at 08:55:58PM +0200, "Mgr. Šimon Tóth" alleged:
>> Semantics of stuff like -l nodes=X,procs=Y should be defined in the
>> documentation, but seriously when ANYONE writes -l nodes=4, he means
>> exactly what he wrote: "I want 4 nodes".
>
> I got it.
>
> We toss "nodes" out the window. Beat it out of the house.
>
> We turn the whole thing around and use "procs" only.  "-l procs=X"  
> makes
> perfect sense. Everyone knows what that means, right?
>
> *AND* "-l procs=X:ppn=Y" makes perfect sense too!  I want X processors
> distributed as Y procs per node.
>
> Noone ever actually cared about nodes anyways. They care about  
> processors. If
> your job is CPU-bound, then all you care about is how many  
> processors you want.
> If your job is IO-bound, then you care about having enough bandwidth  
> available
> to your processor.
>
> We could even go one better and use "cores". How many cores would  
> you like
> today, sir?
>


I like it


>
>> Btw. I still don't think that -l procs=X is a good idea. I would much
>> rather see something like #packed, or #can_pack supported in the  
>> nodespec.
>
> Something like "#packed" could be interesting, as long as we agree  
> that packing
> a job with ppn is dumb, useless, and should be considered an error.
>

I agree here - packed plus ppn doesn't make sense.



>
>
>> Even more great would be disjunctive nodespec support (but that goes
>> into NP-complete even for determining if the nodespec can be  
>> satisfied).
>>
>> "-l nodes=4:ppn=2:vmem=4G+2:ppn=4:vmem=16G|8:ppn=3:vmem=8G"
>
> Sure. Thank you for volunteering :)
>

You're signing up all kinds of volunteers lately