Part of an initial research project was to explore how parallel Kinetic Monte Carlo can be mapped to the architecture of an NVidia GPU.
This code contains my initial experiments to accomplish this. It supports multiple GPUs and, up to my benchmarks at least, it offers a modest speed-up even though it is unoptimized. It is not production-level code; at this point, it can only simulate an 1D Ising system (I also have a version for 2D). It's been a while since I wrote it, so I can't guarantee it will run on your system. I only share it as a proof of concept. I have since moved on to implementing code in SPPARKS; see also [2].
Because it was written with the "fractional-step" point-of-view in mind (see [1,3]), i.e., that an asynchronous parallel kMC algorithm is really defined by the lattice decomposition + computation schedule, I put a lot of focus into making schedules easy to change. This makes the implementation of the Strang splitting a relatively simple addition. I've been writing code to replicate this functionality in SPPARKS.
- Arampatzis, G., Katsoulakis, M.A., Plecháč, P., Taufer, M. and Xu, L., 2012. Hierarchical fractional-step approximations and parallel kinetic Monte Carlo algorithms. Journal of Computational Physics, 231(23), pp.7795-7814.
- Plimpton, S., Battaile, C., Chandross, M., Holm, L., Thompson, A., Tikare, V., Wagner, G., Webb, E., Zhou, X., Cardona, C.G. and Slepoy, A., 2009. Crossing the mesoscale no-man’s land via parallel kinetic Monte Carlo. Sandia Report SAND2009-6226.
- Arampatzis, G., Katsoulakis, M.A. and Plechác, P., 2014. Parallelization, processor communication and error analysis in lattice kinetic Monte Carlo. SIAM Journal on Numerical Analysis, 52(3), pp.1156-1182.