Parallel Computations Reveal Hidden Errors of Commonly Used Random Number Generators

This paper reports that in parallel Monte-Carlo simulations of the 2D Ising-Model, commonly used pseudo-random number generators (PRNG) lead to manifestly erroneous results. When parallel random number sequences for a parallel simulation are generated by a same PRNG with diffrent initial seeds, the sequences can be strongly correlated with each other if the seeds are selected imprudently. The error is due to some dependencies of the outputs of the PRNG on the initial seeds. This type of defect is not so problematic in non-parallel computations, but becomes serious in parallel situations. Special care is required when using such PRNG in parallel computations.

• Type: Conference
• Year: 2006
• Where: CSC
• Authors: Hyo Ashihara, Ai Kuramoto, Isaku Wada, Makoto Matsumoto, Hironori Kikuchi, Masato Kiyama
• Keywords: CSC 2006, CSC 2007, Initial Seeds, Parallel Monte-carlo Simulations, Parallel Random Number