support rank-based OMP places with "vanilla" OpenMP

[naughtont3]

Add support to MOC to use generic (unmodified) OpenMP runtime, but define the OMP_PLACES value as a function of the MPI ranks (i.e., OMPI_PLACES defined in MOC_Init() before OpenMP initialized).

This should be enabled/disabled at compile time to avoid any unexpected interactions when using a modified OMP runtime (e.g., libomp).

NOTE: Focus on OpenMP threads (places) on "cores" use case first. Also, avoid NUMA case at first because we do not have hwloc/fine-grained info.

Target use cases (in order of interest/priority):

• 1.a) 1 rank per physical node -> OpenMP threads (places) on cores
• 2.a) 1 rank per core -> OpenMP threads (places) on cores
• 3.a) 1 rank per NUMA node -> OpenMP threads (places) on cores
• 1.b) 1 rank per physical node -> OpenMP threads (places) on hwthreads
• 2.b) 1 rank per core -> OpenMP threads (places) on hwthreads
• 3.b) 1 rank per NUMA node -> OpenMP threads (places) on hwthreads
• 4.x) Future - 1 rank per GPU -> OpenMP threads (places) ???

[naughtont3]

Added code for defining places in MOC_Init(). Need to revise to have proper syntax for OMP_PLACES definition. Currently get OpenMP error msg like: OMP: Warning #123: Ignoring invalid OS proc ID 11.

Will create WIP PR when that is fixed to keep things linked to this ticket, https://github.com/naughtont3/MOC/tree/tjn-myrank-place-vanilla

[gvallee]

Here are the cases to support, let's keep this issue for overall tracking, I will open individual issues for all these cases.

20180111171506622.pdf

[naughtont3]

Returning to this and did not see the invalid error msg I mentioned above. Successful test was on my Linux desktop. I'll have to try on another machine that has more cores.