Did a web search for the words in the subject line but the first few pages
of suggestions don't appear to address my issue so here goes.

I am seeing jobs requesting say, 4 slots within a PE, having that
request fulfilled in such a way that when running, the slots are
split across two different queues on the same host, both of
which have the PE attached to them.

An anonymised example of a qstat -f -u \*

[email protected] BIP   0/6/16         20.00    lx24-amd64
  12345 0.50281 job_01     user01       r     01/01/1970 14:15:42     1
  12346 0.50245 job_02     user02       r     01/01/1970 17:33:37     4
  12347 0.50198 job_03     user02       r     01/01/1970 15:33:42     1

[email protected] BIP   0/10/16        20.00    lx24-amd64
  12345 0.50281 job_01     user01       r     01/01/1970 14:15:42     3
  12347 0.50198 job_03     user02       r     01/01/1970 15:33:42     7

whilst a slighty different invocation of qstat

qstat -s r -u \* -q queue_02.q@*

shows

job-ID  prior   name       user         state submit/start at
queue                          slots ja-task-ID
-----------------------------------------------------------------------------------------------------------------
  12345 0.50281 job_01     user01       r     01/01/1970 14:15:42
[email protected]     4
  12347 0.50198 job_02     user02       r     01/01/1970 15:33:42
[email protected]     8

suggesting that all of the allocated slots are attached to that queue,
rather, that the starter
process was launched in that queue.

The PE in question here has

allocation_rule    $pe_slots

but I have seen this with fixed number allocation_rule PEs as well.

I appreciate that the slots, as requested are all being allocated on a
single host, as
per the sge_pe man page, but I am surprised, given that there are/were
free slots in
the queue that the first process started in, that they are swilling
about between
other queues.

Any insight and/or ways to prevent the scheduler doing this?

Would assigning a seq_no to each of my queues force things, because
all of the slots
in one queue would, presumably, be "looked at" before the slots in the
next one in the
sequence list were considered?

Kevin Buckley,
ECS, VUW, NZ
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to