Thanks for following up. I have a question to clear things.
1. If I do not specify any affinity in mpirun, what memory allocation policy is used by default?
a. When it is not compiled with libnuma
b. when compiled with libnuma
BTW What do you mean by "internal memory" in OMPI?
Libnuma will be used automatically if you use a paffonity setting.
But note that it only applies to internal memory in OMPI - I don't remember the version offhand, but late in the 1.5 series we set the general affinity policy for all memory (possibly the unreleased 1.5.4?).
Sent from my phone. No type good.
On Jun 30, 2011, at 8:31 PM, "Qasim Ali" <email@example.com> wrote:
> Hi folks,
> I installed open-mpi with the libnuma library using the --with-libnuma. Everything installed fine. When I do
> ./ompi_info |grep maffinity
> MCA maffinity: first_use (MCA v2.0, API v2.0, Component v1.4.2)
> MCA maffinity: libnuma (MCA v2.0, API v2.0, Component v1.4.2)
> I see two policies here. first_use and libnuma
> I have two questions regarding these:
> 1. Which policy will be used in my MPI application?
> 2. Do I have to specify maffinity when using mpirun command or the open-mpi runtime will automatically use the allocation policy (first_use or libnuma?)
> users mailing list
users mailing list