Reuti <[email protected]> writes:

>> Do you mean it takes resource requests and their reservation into
>> account properly?
>
> No, not directly. It scans the output of `qstat` of any given queue
> for waiting jobs (hence: which may run in this queue) and bases the
> decision on this.

I'm sure it's clear to Reuti, but I should have said explicitly that
these things may work reasonably in a uniform environment like that, but
not necessarily on a horribly heterogeneous cluster as we have.
(E.g. waking up nodes in different PEs doesn't help a specific
multi-node job run, likewise one too small for a given single-node job.)

You also don't want to switch things off when they're reserved.

>> Note that power cycling may hurt reliability (see previous discussion
>> here).  Running with the Linux "ondemand" frequency governor
>> significantly reduces consumption of idle nodes without significant
>> problems we've seen.  If necessary (latency effects?) you could flip
>> between performance and ondemand in the usual SGE hooks.
>
> I second this. But I saw locations where the cpufreq daemon wasn't installed 
> and they chose the above solution instead.

Hmm...

-- 
Community Grid Engine:  http://arc.liv.ac.uk/SGE/
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to