OpenMP: Same process on one processor is faster then on two processors

[Reinaldo_Garcia]

We are running code written with OpenMP on a desktop i7 920, 2.66 Ghz processor. We just set up a Dual Xeon 2.53GHz Quad Core server running server 2008 64-bit. The desktop is running Windows 7 64-bit. Running the exact same executable on the server, it takes a minute longer to run than on the desktop. We tested again but this time turning off the second processor on the server, and it runs faster than with two processors enabled.

This table shows the tests we ran.

Computer Speed Processors Threads Runtime

Desktop (i7) 2.66 Ghz 1 8 3 minutes 40 seconds

Server (Dual Xeon) 2.53 Ghz 2 16 5 minutes 53 seconds

Server (Single Xeon) 2.53 Ghz 1 8 4 minutes 18 seconds

Any ideas on why this may be happening or any suggestions would be welcome.

Regards,

R.

[jimdempseyatthecove]

My guess is your working data fits nicely into one L3 cache and the work (data)performed on one thread on one of the processors is passed off to a thread in one of the other processors. (IOW requires additional memory fetch).

Without seeing the actual code it is hard to guess as to the specific cause.

Try setting environment variable

KMP_AFFINITY=compact

Jim Dempsey