On our systems we use:
% qconf -ssconf
...
queue_sort_method load
...
load_formula seq_no*100+m_core-slots
...
% qconf -sq small
...
seq_no 20
...
s_rt 4:00:00
h_rt 4:00:00
...
% qconf -sq medium
...
seq_no 30
...
s_rt 48:00:00
h_rt 48:00:00
...
This causes the system to attempt to use the queue sequence number as
the primary sort order for nodes and then within the queues attempts
to pack the jobs onto individual nodes. Our nodes are divided into
groups with different resource limits and seq_no orders the queues
in the needed order.
The 'm_core-slots' component attempts to keep the larger core count
systems fullest. Right now all our nodes have 8 cores so this hasn't
been verified. We will shortly be adding some new systems with more
cores.
This seems to mostly work as desired, but like lots of things in Grid
Engine other setting and actions can impact what actually happens.
(I think) GE attempts to apply load adjustments when it starts new
jobs so that everything doesn't end up on a single node that was the
lightest load at the start of the scheduling cycle. See sched_conf for
job_load_adjustments and load_adjustment_decay_time.
We are using still SGE 6.2u5.
Stuart Barkley
--
I've never been lost; I was once bewildered for three days, but never lost!
-- Daniel Boone
_______________________________________________
users mailing list
users@gridengine.org
https://gridengine.org/mailman/listinfo/users
