Quality and performance of a pseudo-random number generator in massively parallel plasma particle simulations

Matsuoka, Seikichi* ; Satake, Shinsuke*; Idomura, Yasuhiro   ; Imamura, Toshiyuki*

The quality and performance of a parallel pseudo-random number generator (PRNG), KMATH_RANDOM, are investigated using a Monte Carlo particle simulation code for the plasma transport. The library is based on Mersenne Twister with jump routines and provides a numerical tool which is suitable and easy-to-use on massively parallel supercomputers such as K-computer. The library enables the particle code to increase the parallelization up to several thousand processes without loosing the quality and performance of the PRNG. As a result, the particle code can use large amounts of random numbers, which results in removing unphysical phenomena caused by a numerical noise.