[Omp] OpenMP and Intel compiler

[Gerhard Theurich]

Hi,

If I understand you correctly you are running the pure MPI run with only
2 CPUs while running the hybrid run with 2 x 2 = 4 CPUs? Well, then I
guess what you are seeing is that without debugging flags the execution
is fast enough that the overhead introduced by the hybrid mode is larger
than what you gain from it (this of course should be problem-size
dependent!). Thus you observe that without -g your 2CPU MPI run is
faster than your 4CPU hybrid run. 

Then, when compiling with -g, and effectively slowing down the
calculation (hence both debugging runs take much longer than their
non-debugging counterparts), the overhead introduced by hybridization is
amortized to a good part by the cost of calculation (which now is so
much higher). Consequently you are seeing shorter execution time for the
4CPU hybrid run compared to the 2CPU MPI run. 

If you were to increase the problem size I expect that you will find a
cross-over point also for the runs without debugging mode. However, you
are really comparing 2CPU against 4CPU runs, if I understand you
correctly, so not quite fair.

Gerhard


Laksono Adhianto wrote:
> 
> Hi,
> 
> Thanks for your fast reply (it's 11pm CT, and still lot of people work :-)
> 
> Here is the result of SWEEP3D, 50x50x50 grid with 2 MPI processes (for pure MPI)
> and 2 MPI processes * 2 OpenMP threads (for the hybrid) without optimization
> flags:
> 
> ----- without debugging mode:
> MPI             1.80069398880005
> OpenMP/MPI      10.8658130168915
> ----- with debugging mode:
> MPI             88.6593250036240
> OpenMP/MPI      58.1125251054764
> 
> I run this SWEEP3D twice and the result is the approximately the same. However,
> this is just preliminary test, perhaps it's too early to draw any conclusion. I
> am sure we cannot generalize that using debugging mode in Intel compiler, hybrid
> MPI/OpenMP runs faster than pure MPI :-)
> 
> I just wonder if other people has the same experience....
> 
> Regards,
> 
> ------
> Laksono Adhianto
>