Code optimisation in a nested-sampling algorithm

The speed-up in program running time is investigated for problems of parameter estimation with nested sampling Monte Carlo methods. The example used in this study is to extract a polarisation observable from event-by-event data from meson photoproduction reactions. Various implementations of the basic algorithm were compared, consisting of combinations of single threaded versus multi-threaded, and CPU versus GPU versions. These were implemented in OpenMP and OpenCL. For the application under study, and with the number of events as used in our work, we find that straightforward multi-threaded CPU OpenMP coding gives the best performance; for larger numbers of events, OpenCL on the CPU performs better. The study also shows that there is a “break-even” point of the number of events where the use of GPUs helps performance. GPUs are not found to be generally helpful for this problem, due to the data transfer times, which more than offset the improvement in computation time.