On Mon, Apr 20, 2009 at 5:11 PM, Greg Lindahl <[email protected]> wrote: > On Mon, Apr 20, 2009 at 04:59:31PM -0500, Rahul Nabar wrote: > >> Why would PBS-Torque not allow this and my previous threads >> "JOBNODEMATCHPOLICY EXACTNODE" by default? Are there any reasons not >> to use them? The compromise is not obvious to me. > > In general it's most efficient to have a job use whole nodes. If you > use a partial node, jobs will likely interfere with each other, > reducing overall performance. > > Now if your code only runs on N**2 nodes, using whole nodes can be > painful. But with M*N nodes, you're usually OK.
My code (depending on the specific job at hand ) parallelizes well over a multiple of a small integer. Mostly (a) multiples of 4 or (b) multiples of 9. Since I have 8 cpu/server Job-Type-(a) is great but Job-Type-(b) requires a trade off. If I wanted to request whole nodes I'd have to shoot for 8x9=72 but that degree of parallalization is overkill. At that point the parallalization is no longer so efficient. The ideal situation for our cluster then seems to me: Request full nodes + a fragment. If I can selectively allow these non-full-node-fragments from a specific pool of nodes I could minimize my overall cluster fragmentation. But I've no clue how to implement this. Has anybody tried such a solution? -- Rahul _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
