topic If you want to see whether in Intel® Moderncode for Parallel Architectures

Nested OMP on Xeon Phi using OMP4

james_B_8 — Wed, 19 Nov 2014 21:02:10 GMT

Xeon Phi has 60 cores and 4 threads per core. I am writing an experiment that will have 1 master thread on each core, and each of these will spawn 4 slave threads.

Looking at the manual https://software.intel.com/en-us/node/512835 it seems that I want to set the envars:

MIC_OMP_NESTED=TRUE
MIC_OMP_PROC_BIND="spread, close"
MIC_OMP_NUM_THREADS=60

Is this correct? I've tested this and it doesn't die... Is there a way I can get the runtime to spitout affinity debug info about where it is actually placing things so I can be certain?

Cheers,

James

If you want to see whether

TimP — Thu, 20 Nov 2014 15:01:00 GMT

If you want to see whether spread or close took effect, you could set MIC_KMP_AFFINITY=verbose. Are you serious about wanting to set both at the same time?

One might guess that your intentions align with OpenMP 4 teams concept. OMP_NESTED doesn't have good affinity options.

Thanks.

james_B_8 — Thu, 20 Nov 2014 16:17:29 GMT

Thanks.

KMP_AFFINITY=verbose does indeed output verbose info for OMP_PROC_BIND.

I think it's working with the envars in my OP. Looking at the case of 2 master threads each spawning 2 slaves I get:

OMP: Info #242: KMP_AFFINITY: pid 50593 thread 0 bound to OS proc set {1}
...
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 1 bound to OS proc set {119}
...
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 2 bound to OS proc set {2}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 3 bound to OS proc set {120}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 2 bound to OS proc set {120}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 3 bound to OS proc set {2}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 2 bound to OS proc set {2}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 3 bound to OS proc set {120}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 2 bound to OS proc set {120}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 3 bound to OS proc set {2}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 2 bound to OS proc set {2}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 3 bound to OS proc set {120}
OMP: Info #242: OMP_PROC_BIND: pid 50593 thread 2 bound to OS proc set {120}

0 and 1 are the 'masters' and threads 2 and 3 are the 'slaves' here. It thread numbering for the slaves is a bit tricky here. It seems that they are bound to the OS proc next to the OS proc of their parent. Where they end up depends entirely on which master spawns them first. This seems sensible to me.

I will try out omp teams.

Cheers,

james

James,

jimdempseyatthecove — Sat, 22 Nov 2014 16:20:32 GMT

James,

Try:

MIC_OMP_NESTED=TRUE
MIC_OMP_PROC_BIND="spread, close"
MIC_OMP_NUM_THREADS=60,4

Jim Dempsey