On 02/24/16 12:05, Damian Montaldo wrote:
Hello, sorry if this answer is in the man page but we can't find the
solution
We're working on a cluster that has 64 cores nodes.
Since nodes are not shared but running jobs in exclusive mode, we want
to require and enforce users to ask for a minimum number of cores. We
want to avoid underusing node's throughput.
Is there a way to do so?
Since we are running [GrpCPURunMins] based quotas, when a job locks
the entire node but requires less than the available cores, say n
cores (n<64), it consumes only n * running time [GrpCPURunMins], and
not 64 * running time. So jobs keep 64-n cores idle but get "billed"
only by n...
What is your evidence of this? On my systems it shows just the
opposite, $TotalNumberOfCPUSOnNode * running time no matter what they
request. At least that is what scontrol show assoc_mgr shows.
Thanks in advance,
Damián.
