Accelerating the Computation and Verification of Molecular Collision Models

Presentation Type

Oral and/or Visual Presentation

Presenter Major(s)

Computer Science, Mathematics

Mentor Information

Christian Trefftz, trefftzc@gvsu.edu; Greg Wolffe, wolffe@gvsu.edu

Department

School of Computing and Information Systems

Location

Kirkhof Center 2216

Start Date

13-4-2011 10:00 AM

End Date

13-4-2011 10:30 AM

Keywords

Mathematical Science, Physical Science, Technology

Abstract

Our project constituted a case study in computational science: applying parallel computing techniques to mathematical models for solving a scientific problem. The problem involved a physical chemistry model that evaluated simulations of molecular collision experiments. The collision model was implemented via a 15,000-line FORTRAN-77 simulation. This project was chosen for parallelization because of its extreme computational complexity and significant execution time. We targeted two new and different technologies to parallelize the simulation: OpenMP and CUDA FORTRAN. Nearly linear speedup was measured in the OpenMP parallel version executing on a 16-core multiprocessor. Experimental data indicates speedups should continue to scale well with an increasing number of processors. Results from the CUDA FORTRAN parallel version executing on a graphical processing unit are still pending, but we predict greater speedups will be observed since modern GPUs contain hundreds of stream processors.

This document is currently not available here.

Share

COinS
 
Apr 13th, 10:00 AM Apr 13th, 10:30 AM

Accelerating the Computation and Verification of Molecular Collision Models

Kirkhof Center 2216

Our project constituted a case study in computational science: applying parallel computing techniques to mathematical models for solving a scientific problem. The problem involved a physical chemistry model that evaluated simulations of molecular collision experiments. The collision model was implemented via a 15,000-line FORTRAN-77 simulation. This project was chosen for parallelization because of its extreme computational complexity and significant execution time. We targeted two new and different technologies to parallelize the simulation: OpenMP and CUDA FORTRAN. Nearly linear speedup was measured in the OpenMP parallel version executing on a 16-core multiprocessor. Experimental data indicates speedups should continue to scale well with an increasing number of processors. Results from the CUDA FORTRAN parallel version executing on a graphical processing unit are still pending, but we predict greater speedups will be observed since modern GPUs contain hundreds of stream processors.