Générateur de nombres aléatoires en parallèle

Reliable random number generators are essential for stochastic simulation and for many other applications. In parallel computing environments, we need multiple streams of random numbers that can run independently of each other. This tutorial talk surveys the state of the art on the design and implementation of random number generators for simulation, with an emphasis on the production of multiple streams and substreams of random numbers. Different approaches are discussed, including the classical recurrence-based algorithmic generators, as well as the more recent counter-based generators and dynamically splittable generators, which permit one to create (not only run) the muliple streams in parallel. We also review the basic quality criteria for good random number generators in general, and their theoretical and empirical testing. Finally, we point out issues that would require further study and software that we still miss.