I am currently testing a HP DL575 with four Opteron 6272 processors under Windows R2 Enterprise with 128GB RAM (latest service packs etc).
This server is/was due to replace an older Intel based Xeon machine (single socket, six core) however I have found the performance on the AMD server extremely sluggish. Initially I was testing Matlab alone (simple jobs taking DOUBLE the time to run) so after some research suspected the Intel Math Kernel Library issue. However I put together some other tests using simple scripting (looping a million times with some simple math which doesn't use MKL at all) and i was shocked to see the AMD machine taking almost double the time to execute the same code - compared to the older Intel boxes??
I've run a bunch of benchmarking tools, the only item that sticks out is RAM write speed and RAM latency, which on the AMD box is reported as very slow (latency is about double, write speed is about half, of the other boxes). I've tried swapping out RAM from HP which didn't make any difference. Perhaps the architecture of a quad-socket machine is just different, greater overhead? I have confirmed the RAM setup is as per spec required for maximum performance (with HP).
Any suggestions on what to try? I have tried every BIOS option with no success. Perhaps there is something obvious I am missing?