[torqueusers] Memory resource Limits
jkusznir at gmail.com
Tue Jul 29 12:18:31 MDT 2008
I've been "strongly encouraging" our users to supply memory resource
limits as well as their CPU requests and wall clock limits on our
cluster. However, after some experimentation, we havne't found a
resource limit that performs "as expected".
We are running Rocks 4.2 with torque 2.1.8 and maui 3.2.6.
Our desire is to specify a request that each "process" will be
allocated X ram (and be killed if it exceeds that ram). By
"allocated", I mean, if I say I need 8 nodes with 8GB per process, and
my nodes have 8 cores and 8GB (total per physical node), it will be
smart enough to only allocate one process per physical node. I know I
can work around this with ppn= options and such, but my goal is to let
torque make the "right" decisions and not require the users to know
the underlying archecture and coax torque to map their job how it
should be mapped.
We've tried mem= with no results; pvmem resulted in killing all jobs
unless it was actually using an extremely small amount of ram.
We also have a frequent requirement that one process be given a
different ram allocation than the rest (for example, the first process
is responsible for collecting and disemminating ram and needs 2GB
while all the rest of the processes only need 1GB.
Any suggestions? Is this actually a maui rather than torque problem?
(I still am often confused as to where the line is between torque and
More information about the torqueusers