[Beowulf] Question: Who has apps that regularly swap?

[Sean Dilda]

On Tue, 2004-08-31 at 00:41, Mark Hahn wrote:

> actually, that leads to another survey question: do you schedule
> your job by loadaverage or according to a resource-allocation scheme?
> when a job runs on our systems, it's pretty much guaranteed a 
> free CPU.  this actually causes some strife with people who want 
> to run non-compute-bound jobs, or jobs which have dynamic amounts
> of parallelism...

For my cluster its somewhat of both.  We use SGE which by default will
schedule the job on the machine with the lowest load average, with some
special allocation rules used for parallel jobs.  In addition, users can
request certain resources (such as so much free memory), and SGE will
then only schedule jobs on nodes which meet your requirements (and the
nodes with the lowest load avg of the ones that do).  Or you can make
the request a "soft" one where it will try to fullfill your request, and
if it can't, it will then schedule the job on a node that doesn't meet
it.  All in all I think it works pretty well once you've twisted all the
knobs the right way.


Sean