Testbed Support for GridPP member institutes
> [mailto:[log in to unmask]] On Behalf Of John Gordon said:
> Are there any 'assumptions made about how logical CPU maps to
> "number of allowed jobs"'. I mean in YAIM or the BDII info
> providers, not in someone's head.
As Steve said, in the info system they're separate, the dynamic state is
in the CE. The connection is an indirect one via APEL, i.e. over some
period of time you can see what resources APEL thinks were consumed by
jobs. If that turns out as more than the published total capacity then
there would clearly be a mistake somewhere. If it's less it's fairly
normal because you have downtime and the system is usually not full, but
still as we move into data taking with more contention I assume that the
MB will be looking for sites which appear to deliver a long way below
their claimed capacity.
> Logical cpus are used to calculate the total power so if you
> limit the number of jobslots you should probably only run the
> benchmark with that number of copies.
Yes.
> This sounds rather
> convoluted. If you are maxing out on network or disk perhaps
> you shouldn't enable hyperthreading?
That's really a decision for sites about how to configure things to get
the best use of their system.
Stephen
--
Scanned by iCritical.
|