NUMERICAL REPRODUCIBILITY, PORTABILITY AND PERFORMANCE OF MODERN PSEUDO RANDOM NUMBER GENERATORS: PRELIMINARY STUDY FOR PARALLEL STOCHASTIC SIMULATIONS USING HYBRID XEON PHI COMPUTING PROCESSORS

摘要

Stochastic simulation is of key importance for scientific computing. Parallel stochastic simulations are becoming popular since the Monte Carlo method is easily parallelized. However, the main interests we observe in related publications deal with the following issues: simulation performance, applicability of the simulation, model design, etc. Even if these interests are justified, accuracy and reproducibility, especially for numerical simulation experiments, must remain a major concern in the process of increasing and diffusing scientific knowledge. We recently realized that this essential question of reproducibility is sometimes overlooked. In this paper we study the essence of reproducibility for stochastic simulations by focusing on some of the best random number generators because they are at the heart of stochastic simulations. We have identified some reproducibility problems linked to different compilers, operating systems and also with a virtual machine. In this paper, we check the portability and performances of modern pseudo-random number generators that are tested on hybrid computing nodes with the latest Xeon Phi processors.