LHC Computer Grid - Rollout
> [mailto:[log in to unmask]] On Behalf Of Jeff Templon
said:
> gotta agree with the traylenator here.
Maybe we should be clearer about what we're agreeing or disagreeing
about ... I'm not suggesting that time limits are irrelevant, but that
they're *easy*, at least relatively - the glue schema already represents
them as far as the job selection goes, so all you have to do is the
fairly simple step of passing them through to the qsub. And because
they're a policy they don't interact with the choice of which WN to run
the job on. Other resource constraints are quite a bit harder, at least
if your nodes are heterogeneous. And bear in mind that they get even
harder once you have multiple constraints - you may have nodes with 4 Gb
of RAM, and you may have nodes with 500 Gb of free disk space, but you
can't necessarily satisfy both at once ... (again this doesn't apply to
the time, you can have any time limit with any WN). For your box analogy
you need to know all the dimensions, being able to fit each one
separately isn't good enough. (Although I seem to recall that the
British post office used to limit parcels to something like three feet
as a linear dimension - so if you wanted to send a five-foot fishing rod
you had to put it along the diagonal of a three-foot cube :)
Stephen
|