I'm not an IT guy, but an engineer using the software and computers for engineer’s applications namely CFD (Computational Fluid Dynamics).
The whole topic is not easy to describe so I would try to give as much information as possible in the shortest/simplest way.
To be able to run our calculations in parallel we use Mpich 2 environment. Then while running the application in parallel way I can divide any case into processes and than subscribe each process to a processor(s). I can load each processor as much as I wish (at some extend which BTW is not that simple but possible). This means that having a case divided to 4 processes I can have 4 no (or less) of cores or threads (using HT) to be specifically designated with (ideally) up to 100% of core load/use for each core/thread.
Testing different computer’s configurations our finding is that in our case not the frequency of a processor is the most important factor but the frequency of the ram memory. The testing platform was a computer with i7 x980 (HT on), 24 GB RAM DDR3 and Asus P6T delux v2 MB, where we could check the calculation time for different configurations, such as CPU 3.33 - 4.4 GHz and RAM 667 – 800 GHz. Our finding was that the RAM frequency was a deciding factor to shorter (quicker) the calculation time not the CPU clock. With higher RAM frequency we could get 20% improvement over 5% for max CPU clock.
We have possibility to use a platform build of 2 x X5680 and some platforms build with i7 six core family. But we think about going to a machine build on E7 2800/4800/8800 which in configuration of two processors can give up to 20 cores (40 threads with HT on). As the processors are really expensive we can’t basically afford buying it to check how it goes (unsuccessful test purpose). The difficulty is that there are no real benchmarks which can really represent the difference for parallel calculations in mpich2 environment and our applications. For example the small increase of performance between CPU with 3.33 and 4.4 GHz frequency (5-10%) in terms of calculation speed. While talking about calculations time I mean days not hours.
So which of those two (or maybe some other processors) is better for parallel calculations?
Which from the E7 family would be optimal for our use as it is difficult to judge on the spec table?
Is there any way to test the processor or configuration prior buying (even remotly) just to see how faster the set up is?
We would like to build a two processors platform, so a tower rather a rack system would be preferable.
thx in advance for answers