Am 12.05.2008 um 18:01 schrieb Craig Tierney:
Reuti wrote:
Hiho,
Am 12.05.2008 um 15:14 schrieb Prentice Bisbal:
It's still an RFE in SGE to get any arbitrary combination of
resources, e.g. you need for one job 1 host with big I/O, 2
with huge memory and 3 "standard" type of nodes you could
request in Torque:
-l nodes=1:big_io+2:mem+3:standard
(Although this syntax has its own pitfalls: -l nodes=4:ppn=1 might
still allocate 2 or more slots on a node AFAIO in my tests.)
You mean the syntax has its pitfalls in Torque,
How Torque implement it for now: With ppn=1 I want one core per node,
but might end up with any other allocation. AFACS they don't have an
allocation rule like SGE where you can put a fixed 1, 2, $round_robin
et al. there.
or how SGE may impelement
it?
If it would be in SGE, it would be a point of discussion how to
interpret this expression.
I personally like the way SGE allocates nodes. I can control how
they get nodes. When a user asks for 16 processors (core, slots,
whatever)
they should get N nodes that have M processors, and N*M=16. If a user
needs to specify ppn=2 (or 4 or 8) it means they will mess it up
causing
jobs to share nodes and adversely impact each other which I don't
want.
As ppn=2 will allocate 2 slots on a node (and avoid further usage),
it shouldn't interfere with other users. I never saw a problem with it.
But with an old Linda version we had the problem (<= Gaussian 03 C.
02), that you had to specify the nodes and for all nodes you need in
addition to specify how many slots to use: 1 or 2 or 4 slots - and it
must be the same on all nodes. If a node was double in the list,
Linda complaint. So the only option was to specify full nodes: "-l
nodes=2:ppn=4" to get the complete node and sometimes to wait,
although an allocation 2+2+2+2 was possible and free in the cluster.
But requesting "-l nodes=4:ppn=2" could end up with an allocation 4+2+2.
-- Reuti
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
