I am running some muti-threaded code that uses a lot of memory (~6GB). I have it compiled on a cluster that uses Xeon 5160 quad core CPUs with 4MB cache/core, 8GB of RAM and kernel 2.6.18 (x86_64) and I have also the same code compiled on another cluster that uses Opteron 2214 quad core CPUs with 1MB cache/core, 8GB of RAM and kernel 2.6.28 (x86_64). On the Intel cluster the code generates barely any system load when I look on top but on the AMD cluster it can generate over 50% of system load. On the intel machines my process can generate a constant load of 3, while the load generated by the process on the AMD machines is about 1.6.

Does anybody know what could cause this? Could it be the version of the kernel or the difference between the CPUs (for example, the much smaller cache of the AMD CPUS)? My software and all its dependencies are the same on both clusters. I am also using the exact same version of GCC to compile everything.

Thanks!

What could produce a big difference in system load between these two systems? Is that possible that these AMD CPUs have poor multi-threading performances with pthread?

My jobs run about 5 times slower with the AMD machines... It seems that half of this is due to the CPU frequency/cache and another half is due to the generated system load.