[GE users] One slot per host, or a way to consume all slots
reuti at staff.uni-marburg.de
Sun Feb 25 14:03:37 GMT 2007
Am 25.02.2007 um 03:49 schrieb FL:
> You mean you want the complete machine for the job because of memory
> And because of processor constraints: we have single P4s
Aha, now things come clear.
> This formula I don't get: why are you using (workerjobs-1)? If you
> refer to Lindaworkers, a parallel Gaussian job with a proper %
> lindaworker list and 4 cpu requested would also need "-pe linda 4".
> The question of requesting the number of CPUs was determined by
> observing that if your gaussian-Linda .com script has
> %nprocl = n
> then use
> #$ -pe gauss nprocl-1
> otherwise you get nprocl+1 tasks (a master + nprocl tasks).
I remember such a strange behavior in G98-A.11 and AFAIR it was fixed
in A.13. Are you still using G98? In G03 I see a correct behavior.
If you mean the output of "qstat -g t": it reflects only the granted
slots, and whether you get an additonal slave-task listed there
depends on the setting of "job_is_first_task" which is TRUE for
> If I had a cluster like that, then the users wouldn't be so concerned
> about unrelated jobs running on the same node. But I only have single
> (core) P4s with 1GB memory. (OK I have 6 core duos with 4 GB, but
> this is another story--I would want the job allocation scheme to be
> consistent: attempt to assign one Linda worker per core, and
> exclude anything else).
> Perhaps my understanding of the queing system is wrong: if I have
> two queues A and B which can assign one slot each to host H (both
> queues have host H in their host list) then the number of slots
> to H is 2 (the sum of the slots per host assigned to the queues) and
> so I could get two jobs running on the same host H if they were
> from different queues.
Correct. But this you can limit of you set slots=1 also in the
complex_values definitions of each of your exec hosts.
> Perhaps what I need to do is set the number of slots on each
> host to 1, even though that will starve non-gaussian jobs that
> don't need
> a whole P4...(or a whole core).
What do you mean by "...don't need a whole P4"? If you run two
programs at once on the same machine they will share the CPU time,
hence run slower.
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net
More information about the gridengine-users