SBIR-STTR Award

Highly scalable parallel random number generators (RNGs) will be developed, evaluated and implemented for use in high performance computing on thousands of multi-core processors and general purpose graphics processing units. The main contributions are: (a) design and implementation of new parallel test methods that capture the inter-stream correlations exhibited in practice and complement the currently widely used sequential test batteries, (b) development of new parallel RNGs that produce 100s of thousands of high quality individual random number streams and explicitly minimize inter-stream correlations, and (c) preliminary implementation of the new test methods and parallel RNGs. The proposed RNGs will also be evaluated and tuned to generate cryptographically-secure random number streams that resist cryptanalysis attacks by insiders and eavesdroppers when used in large-scale peer-to-peer and distributed security applications. The investigators have extensive experience in the applications of random number generators, the test methods for random number generators, and the implementation of Monte Carlo applications on large clusters of processors and graphics processing units. The proposed approach balances the theoretical research with the implementation efficiencies and the use in real applications.

Keywords:
Random Number Generators, Parallel Statistical Tests, Inter-Stream Correlations, Graphics Processing Units, Multi-Core Processors, Massively Parallel Computing.

A parallel random number generation suite that generates, dynamically, millions of parallel pseduorandom number streams suitable for consumption by high performance computing (HPC) applications on various platforms will be implemented, tested, and provided to Government, academic and private HPC users for their evaluation and use. The software will be tuned for scalability, repeatability and efficient generation of random numbers for applications running on clusters of multicore processors as well as graphics accelerators. The context-aware random number generators that can automatically provide distinct random number streams based on program context to improve the quality of random numbers consumed by applications will be implemented. A comprehensive test package that incorporates a novel parallel statistical test method for direct estimation of correlations among parallel random number streams will be implemented. This package will have a flexible and adaptable interface and be capable of online testing of parallel streams, allowing users to test any new parallel random number generator, use customized test methods, and assess the quality of the random numbers consumed in an application execution. The investigators have extensive technical expertise and experience in the theoretical and implementation aspects of parallel random generators, parallel architectures and computing, and the test methods.

Keywords:
Random Number Generation, Parallel Statistical Tests, Interstream Correlations, Multicore Processors