Uniform Random Number Generators for Supercomputers

We consider the requirements for uniform pseudo-random number generators on modern vector and parallel supercomputers, consider the pros and cons of various classes of methods, and outline what is currently available. We propose a class of random number generators which have good statistical properties and can be implemented efficiently on vector processors and parallel machines. A good method for initialization of these generators is described, and an implementation on a Fujitsu VP 2200/10 vector processor is discussed. 1 Introduction -- Requirements Pseudo-random numbers have been used in Monte Carlo calculations [1, 3, 15] since the pioneering days of Von Neumann [26]. With the increasing speed of vector processors and parallel computers, considerable attention must be paid to the quality of random number generators available in subroutine libraries. A program running on a supercomputer might use 10 8 random numbers per second over a period of many hours (or months in the case o..