Accelerate the simulation with CUDA

Hi Ocean,

using GPU computing (CUDA / OpenCL) to accelerate ESyS-Particle is an idea which has been discussed between the developers several times. The result of the discussions was that it would be nice to have a look at the issue, but that we didn't have the neccessary resources.
The problem is hard for a couple of reasons:
- due to the structure of the code significant changes would be required to make it suitable for GPU computing. One of the key reasons is that ESyS-Particle calculates the inter-particle forces by walking through a list of particle-particle interactions and computing their forces. This is a lot less suitable for vectorisation (which GPU computing relies on to get decent performance).
- another issue is that ESyS-Particle does need "double" precision calculations, i.e. single precision isn't sufficient. This unfortunately rules out most affordable GPU hardware. In particular recent "consumer level" NVidia GPUs have atrocious DP performance.

To answer your specific questions:
1) In principle it should be possible - DEM on GPUs has been done before. However, the work required would be significant (PhD-project?)
2) If this problem is tackled, the unix-version is probably more suitable
3) Speeding up the contact detection only is a) rather tricky due to the parallelization / domain decomposition involved and b) probably doesn't gain much performance because most of the time is spent in the force calculation, not in the contact detection, in particular in small displacement problems like uniaxial compression tests.

Additional question: on your i7, how many worker processes did you use? I'm asking because the i7-2600 has only 4 physical cores, the 8 "virtual" cores are obtained by hyperthreading. Therefore I would expect that you probably loose more performance to to parallelization overhead than what you gain through better CPU utilization when using more than 4 processes (3 workers + master) .

Steffen