Random Number Generation for High Performance Computing

Frontier Technology, Inc. and University of Rhode Island Physics department propose to develop innovative, scalable random number generators for use on multiple parallel computing architectures. Our Phase I effort will include a comprehensive assessment of currently available algorithms for parallel random number generation as well as the currently available tests designed to uncover statistical defects. The considerable experience of the researchers involved in this proposal strongly suggests that the simplest and most versatile method for constructing new, more scalable algorithms will be to work with combinations of some of the most reliable generators currently available and that application based tests which rely on comparison with exactly solved models of continuous phase transitions will be among the most reliable. Random number generator algorithms and hardware specific implementations will be analyzed and optimized given the computational constraints. The most reliable and distributable of these algorithms will be implemented on multi-core systems, parallel computing clusters, graphics processing units (GPUs), GPU clusters and other parallel architectures. These algorithms will be compiled into a standardized library for release as part of a Phase II effort. A prototype of this library will be made available upon the completion of the Phase I effort.