Pseudo-Random Number Generators for Massively Parallel Discrete-Event Simulation

Abstract

A signi cant problem faced by scienti c investigation of complex modern systems is that credible simulation studies of such systems on single computers can frequently not be nished in a feasible time. Discrete-event simulation of dynamic stochastic systems, allowing multiple replications in parallel (MRIP) to speed up simulation time, has become one of the most popular paradigms of investigation in many areas of science and engineering. One of the general problems related with distributed simulation is the need of parallel generation of multiple sequences of pseudo-random numbers across cooperating processors, with the number of known, good paral- lel generators being very limited. This report assesses currently known techniques proposed for generation of pseudo-random numbers in processing systems, particularly the statistical proper- ties of multiple sequences of numbers generated in parallel, and the speed of generation of these parallel streams and also the pseudo-random numbers themselves. Parallel implementations of the MRG32k3a and DX-120-2 generators are found to be the most suitable of those tested.