Benchmarking entry-level HPC machines

From Theory of Measurements Wiki

Available in PDF.

We now got our first p5 510 machine, and here are some preliminary results:

Stream is slightly faster than on the machine that I tested (we have two more memory sticks, and still the slots are half empty).

Function      Rate (MB/s)   Avg time     Min time     Max time
Copy:        6908.4686       0.0046       0.0046       0.0047
Scale:       7221.8309       0.0044       0.0044       0.0045
Add:         5797.9090       0.0083       0.0083       0.0083
Triad:       5604.8606       0.0086       0.0086       0.0086

Here are the results on the machine with two 1GB memory sticks (Add and Triad are faster for some reason):

Copy:        6192.4970       0.0207       0.0207       0.0207
Scale:       5967.9511       0.0215       0.0214       0.0215
Add:         6196.5710       0.0310       0.0310       0.0310
Triad:       6273.5177       0.0306       0.0306       0.0307

Asteroid simulation test, to see speed difference between Power5+ and G5 in the observatory. The speed difference is most likely due to a better implementation of transcendental functions on the Power5+ machine.

Machine      compiler command   real          user       sys
Power5+      xlf -O5            1m17.22s      0m57.74s   0m0.02s
G5           xlf -O5            6m40.358s     6m16.310s  0m0.670s