Parallel algorithms for the Gibbs ensemble Monte Carlo simulations

Conventional Gibbs ensemble Monte Carlo simulations have been carried out on serial machines with the samples of both the boxes being on the same processor. With the advent of distributed memory parallel supercomputers, it is necessary to develop new techniques to parallelize these simulations. In this work different procedures for the parallelization of the GEMC simulations have been tested in an attempt to find the most accurate and efficient method. Since the GEMC simulations involve two simulation boxes which are almost independent of each other except for the time when the exchanges are taking place, we have experimented with techniques which involves placing the two boxes on different processors. Since the boxes are now on different nodes, they need to get paired before any transfers can be made. The different possibilities in which the nodes can get paired lead to a number of different procedures for carrying out the simulations. It is seen that the parallel procedures are more efficient because they can sample more configurations in the same wall clock time, thus improving the statistical accuracy of the results. For the multi-processor codes, it is seen that the swapping of the simulation boxes after a certain number of cycles can lead to a faster equilibration of the processors. However swapping is effective only at lower temperatures when the convergence rates are slow.