Strange Upwelling Performance on Different Architectures

[ejens005]

Hello, I just tested upwelling (test case) on 3 different architectures, and the results are not what I expected.

I used OpenMP, Intel compilers. I installed zlib, hdf5, m4, netcdf, and netcdff.

(1) Intel(R) Xeon(R) CPU E5-2650 v1 @ 2.00GHz, 16 cores. 4*4 grid: wall time 17s

(2) Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz, 16 cores. 4*4 grid: wall time 27s

(3) Intel(R) Xeon Phi(TM) CPU 7210 @ 1.30GHz, 64 cores (KNL). 8*8 grid, wall time 28s


Any idea why the oldest, slowest chip could be fastest, and by such a wide margin? Thanks.

[kate]

The UPWELLING case is not large enough to be a good test of scalability. You are instead testing the interconnect between the cores. You might also learn something by running it in serial mode on all three systems. We designed the BENCHMARK case for these tests - it comes in three sizes so you can time it across that dimension as well.

The place I used to work at once got two new systems at once. According to the BENCHMARK case without saving netcdf files, system A should have been faster. In practice, the output on it was so slow system B was faster for my realistic applications. There's more than one way to slow your code down...

[shchepet]

.....16 cores. 4*4 grid: wall time 17s
.... 16 cores. 4*4 grid: wall time 27s
.... 64 cores (KNL). 8*8 grid, wall time 28s

Am I interpret correctly that "4*4 grid" means tiling 4x4? If yes, read
viewtopic.php?f=17&t=2001&p=7771
-- used to be hot topic sometime in the past.

Still relevant today, see
http://www.cesm.ucar.edu/events/worksho ... petkin.pdf
and scroll down to
Poor Man’s Computing: Overlooked and underutilized resources
then look at table.