[glue-wg] Minutes of today telecons + reminder tomorrow telecon3 PM GMT

[Burke, S (Stephen)]

glue-wg-bounces at ogf.org [mailto:glue-wg-bounces at ogf.org] On Behalf Of
Gerson Galang said:
> As what we've talked about before in the telecon on 22 January, we
need to have a new term to describe the free CPUs for the MPI job,
FreeParallelCPUs perhaps?.

How does this work? I would have assumed that if a system allows MPI it
would allow it on any CPUs. Are you saying that if a system has, say, 16
free CPUs and will allow up to 12-processor MPI jobs it may still be the
case that a 12-processor job won't run because only 10 of the free CPUs
are free for parallel jobs?

>A comment on Stephen Burke's comment about GLUE 1.3 Issues/Comments #3:
>We don't think there's a way to calculate WorstResponseTime on a
preemptive system. A worst response time for a job that user A submits
now might not be valid anymore 5 minutes later because user B who has a
higher priority submits a job and preempts the job submitted by user A.

I still think this is a misunderstanding. Even on non-pre-emptive
systems the values may be wrong in 5 minutes time because more jobs may
be queued in that time and they may go in the queue in front of yours,
the published values are only ever a current snapshot. (Conversely, jobs
in front may be cancelled or executing jobs may abort so the times may
also suddenly get shorter.)

Stephen