Random Number Generation On Parallel Processors

Recent development of high speed supercomputers has enabled us to perform large scale Monte Carlo simulations which need a tremendous amount of random numbers. There are two types of supercomputers, i.e. pipeline type and processor array type, and we will confine ourselves in this paper to the randon number generation on a computer of the latter type. It is desired that not only a sequence of random numbers generated on each processor is a good quality but also sequences generated on different processors are uncorrelated. If we use a linear congruential method on a 32-bit supercomputer, the whole period may be consumed in several seconds to several minutes. Random numbers with a much longer period can be generated by a GFSR algorithm. Using this algorithm, we will propose a method to generate uncorrelated series of random numbers on parallel processors.