[torqueusers] question on memory allocation
srash at yahoo-inc.com
Thu Oct 26 14:38:14 MDT 2006
There is a flag or param about resources..
Dedicated, utilized, or both
(former being what you say it'll take, latter what it's actually using)
and another around that I think
This reminds me of a request wrt torque and enforcing resource requirements.
Would a future feature be to at least globally, if not per resource
(walltime, memory, etc) say that when I say a job has needs 1.4gb of mem and
a walltime of 2 days, that the 1.4gb is entirely soft meaning just use it in
scheduling (ie, each node has 3gb, so only 2 jobs of this type should go
there), but enforce walltime of 2 days (no way on earth our jobs should take
more than that, more like 12 hours max).
Last we (Garrick) and I talked about this, the code for enforcing these
resource 'limits' was only enabled if certain ones were defined. Memory did
not enable it, but walltime did. So we currently only specify walltime, but
sometimes get hung jobs due to transient issues and would love to set a max
walltime of like 12-24 hrs, but if a process w/1.4gb mem setting hits 2.5,
let it swap...(but still don't shoot yourself in the foot by knowingly
putting 3 x 1.4gb procs on a 3 or 4 proc box)
Has this changed since 2.1.2? Any plans?
srash at yahoo-inc.com
From: torqueusers-bounces at supercluster.org
[mailto:torqueusers-bounces at supercluster.org] On Behalf Of Lippert, Kenneth
Sent: Thursday, October 26, 2006 11:24 AM
To: torqueusers at supercluster.org
Subject: [torqueusers] question on memory allocation
I put memory requirements ala' "-l mem=1gb" in all of my job scripts.
When Torque/Maui is deciding where to send the job, does it look at the
ACTUAL free physical memory available on the nodes, or does it look at
the memory on each node minus the sum of the "mem" values of the jobs
already running on that node ? (all the nodes have several np's, from 2
If my memory estimates were exact it wouldn't matter, but they aren't.
I have to estimate a little high, so if the algorithm is the latter,
Torque could be thinking a node was "full" when it really wasn't.
Looking at the documentation, I am thinking it is the first way, but I
want to be sure.
Thanks very much.
torqueusers mailing list
torqueusers at supercluster.org
More information about the torqueusers