[torqueusers] CPU affinity for SMP nodes
Ronny T. Lampert
telecaadmin at uni.de
Thu Apr 27 08:20:04 MDT 2006
> It is just suggestion, but it would be nice to have implemented in
> Torque mechanism to control CPU affinity on shared memory nodes. This
This is a nice suggestion. I had that while ago, too.
But when you think more thoroughly it is not that good (at least these are
my conclusions that prevented me from patching affinity in):
1) The kernel should know best about what process belongs where because only
it knows the "whole picture".
It is not that easy anymore with irq-balancing (the process causing the IRQs
will be more likely scheduled on the CPU handling them) to simply say: put 1
process on 1 CPU and never migrate.
Recent Linux schedulers are very good a figuring all that out, including
cache hot times, load balancing and interactive VS I/O VS compute processes.
2) What to do when there are more jobs than CPUs to be handled?
3) What about "interactive" nodes/jobs?
4) What about jobs that are not strictly 100% CPU & no 0% I/O ones?
5) Working around the CPU scheduler of a bad OS -> better change OS.
More information about the torqueusers