mpi data limit size for speed-up


I implement mpi Fortran code in Intel® Core™ i7-6700K Processor
(8M Cache, up to 4.20 GHz) hardware …

I decompose the data per-processor for my algorithm …

for small data size (grid size) I get speed-up about 4 (8 threads to 1 thread)

when I increase the data size , the speed-up is decreasing…

is it have a reason that when I increase data per-processor the speed-up is smaller ?

Hi @@ali@@,

You’ll need to perform an analysis to determine the cause. The PGI profiler, PGPROF, is MPI aware, so I would suggest to start by profiling your code to see where the differences are. See:

Do you have two i7-6700K chips in your system? If not, this means that you are oversubscribing your cores by using hyper-threads and this can lead to slowdowns. Your chip only has 4 physical cores with two hyperthreads per core. Hyperthreads are good for throughput but not for speed as the two threads will be competing for the same resources on the core.

  • Mat

@mkcolg thanks for your reply …

no… I have one chip … decreasing in speedup is occurring even for number of threads equal to 4