Hi Daniel, I believe you didn't understand what Darryl wrote. It's probably just a language issue. The sentence below says "not just help", not "not help". What he means is that in addition to helping with locality issues, it also helps with thread migrations.
So you are right (and that's what Darryl says as well). One additional caveat and that is that MPO is advisory only. It will allocate memory in the local lgroup if it's available, but will allocate from another lgroup if it's not. So if that lgroup is oversubscribed, you will get remote memory. Regards, Dave Miller Daniel, Wu wrote:
Some expert said it doesn't help. http://blogs.sun.com/d/entry/numa_binding_and_openmp it said "It is worth pointing out that binding does not just help memory locality issues." but if a process was bind to a cpu and never migrate to other cpu, MPO only allocate memory to the home lgroup, so I think the memory used will be always local, and should help. my understanding is wrong?
_______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
