Vectorized Monte Carlo
Thesis/Dissertation
·
OSTI ID:5226790
Examination of the global algorithms and local kernels of conventional general-purpose Monte Carlo codes shows that multigroup Monte Carlo methods have sufficient structure to permit efficient vectorization. A structured multigroup Monte Carlo algorithm for vector computers is developed in which many particle events are treated at once on a cell-by-cell basis. Vectorization of kernels for tracking and variance reduction is described, and a new method for discrete sampling is developed to facilitate the vectorization of collision analysis. To demonstrate the potential of the new method, a vectorized Monte Carlo code for multigroup radiation transport analysis was developed. This code incorporates many features of conventional general-purpose production codes, including general geometry, splitting and Russian roulette, survival biasing, variance estimation via batching, a number of cutoffs, and generalized tallies of collision, tracklength, and surface crossing estimators with response functions. Predictions of vectorized performance characteristics for the CYBER-205 were made using emulated coding and a dynamic model of vector instruction timing. Computation rates were examined for a variety of test problems to determine sensitivities to batch size and vector lengths. Significant speedups are predicted for even a few hundred particles per batch, and asymptotic speedups by about 40 over equivalent Amdahl 470V/8 scalar codes arepredicted for a few thousand particles per batch. The principal conclusion is that vectorization of a general-purpose multigroup Monte Carlo code is well worth the significant effort required for stylized coding and major algorithmic changes.
- Research Organization:
- Michigan Univ., Ann Arbor (USA)
- OSTI ID:
- 5226790
- Country of Publication:
- United States
- Language:
- English
Similar Records
Vectorization methods development for a new version of the KENO-V. a criticality safety code
Computational efficiency using the CYBER-205 computer for the PACER Monte Carlo Program
Investigation of pattern recognition techniques for the indentification of splitting surfaces in Monte Carlo particle transport calculations
Journal Article
·
Mon Feb 28 23:00:00 EST 1994
· Nuclear Science and Engineering; (United States)
·
OSTI ID:5101424
Computational efficiency using the CYBER-205 computer for the PACER Monte Carlo Program
Conference
·
Sun Sep 01 00:00:00 EDT 1985
·
OSTI ID:5342570
Investigation of pattern recognition techniques for the indentification of splitting surfaces in Monte Carlo particle transport calculations
Technical Report
·
Fri Aug 01 00:00:00 EDT 1975
·
OSTI ID:4157716