David Akin wrote:
All,
I'm trying to get the OpenMP portion of the code below to run
multicore on a couple of 8 core nodes.
I was gone last week and am trying to catch up on e-mail. This thread
was a little intriguing.
I agree with Ralph and Terry:
*) OMPI should not be binding by default.
*) There is nothing in your program that would induce binding nor
anything in your reported output that indicates binding is occurring.
So, any possibility that your use of taskset or top is misleading? Did
you ever try running with --report-bindings as Terry suggested?
The thread also discussed OMPI's inability to control the binding
behavior of individual threads. You can't manage individual threads
with OMPI; you'd have to use a thread-specific mechanism, and many OMP
implementations support such mechanisms. The best you could do with
OMPI would be to unbind or bind broadly (e.g., to an entire socket), and
that policy would be applied to all the threads within the process.
But, all that should be unnecessary... there shouldn't be any binding by
default in the first place. I'd check into whether these threads really
are being bound and, if so, why.
