Quasirandom Number Generators for Parallel Monte Carlo Algorithms

A method for generating sequences of quasirandom numbers allows conventional serial Monte Carlo algorithms to be parallelized using a leapfrog scheme. Specifically, a Sobol' sequence can be broken up into interleaved subsets; with each processing node calculating a unique subset of the full sequence, all of the computational advantages of quasirandom Monte Carlo methods over pseudorandom algorithms or grid-based techniques are retained. Tests with several parallel supercomputers demonstrate that as many as 106integration points (up to 6 dimensions) can be generated per second per node in the optimal case where the number of nodes is a power of 2. The speed of communication-free parallel Sobol' sequence generators and the rapid convergence properties of quasirandom Monte Carlo schemes indicate that the method described here may be gainfully applied to a wide range of problems.