Hii Morris, many thanks for the reply.. But My question is different.
I am fine with memory usage..
When a user submit his job with more memory demand, let say with all
memory with on a node and 1 cpu core, sacct calculates the
cputimeraw for only one core. In fact since all the memory is deticated
to the job, the idle cores on that node are not usable.. So I think
there may be another statistics like cputimeraw but takes the idle but
not usable cores into account due to the jobs other kind of demands
(gres, memory etc..)
Regards..
On 12/25/2014 06:20 PM, Morris Jette wrote:
See MaxMemPerCpu
On December 25, 2014 5:30:41 AM PST, Sefa Arslan
<[email protected]> wrote:
Hi ,
I cant find a proper name for the thing I described below. maybe
overAllocatetime??
we have compute nodes with 16cores and 256 GB memory. By default user
can use 16GB memory per core max... But when he/she send the job via
--mem parameter, practically more memory than 16GB can be used per core.
it is fine, some jobs need more memory...
for example a job sent to queue with N1 n1 --mem=256GB, the
cpurawtime, cputime is calculated for only one core (by squeue or
sacct), in fact the full node is acclocated for the job since the all
memory requested on job submission. So no other job run on this node
until the job is end. For billing or just for usage statistics, the
usage should be considered not only according to the number of cpus
requested, but also total allocated system source should be considered.
Is there a term for such kind of usage?
totalCpuAllocateTime or
totalNodeTime etc.
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
