Danny,
I believe one of our customers would really support this change. I'm no
cpu binding expert but it seems to be your proposed behavior should be
better than the current behavior in most situations.
Thanks,
David
On 05/20/2014 05:26 PM, Danny Auble wrote:
The current task/affinity when cyclically binding tasks using more
than 1 cpu will bind to cpus cyclically as well.
We feel this multi-cpu task should be bound in more of a block method,
so a task is bound to close cpus instead of spread potentially across
sockets.
In example of a 2 socket 6 core per socket system, a request of -n2
-c2 would result in the following binding...
task 0 : socket 0 core 0 and socket 1 core 0
task 1 : socket 0 core 1 and socket 1 core 1
Our idea would be to change this default behavior to have the
following binding...
task 0 : socket 0 cores 0-1
task 1 : socket 1 cores 0-1
We will probably introduce a new distribution method so if someone
wanted to get the full cyclic distribution they could.
Does anyone have any feelings on this matter one way or the other?
Thanks,
Danny
