Load balancing and performance issues for the data parallel simulation of stiff chemical nonequilibrium flows

摘要

A new data parallel program for solving stiff chemical nonequilibrium flows on the Connection Machine CM-2/200 and CM-5/5E is presented. The program is written in CM Fortran and employs a timestep-splitting algorithm to solve the governing equations for an unsteady, compressible reacting flow on a structured Cartesian grid. An explicit high-order monotone algorithm with nonlinear damping is used to integrate the convection terms, and a hybrid asymptotic/modified-Euler approach is used to solve the system of ordinary differential equations from the chemical source terms. Integration of the fluid dynamics was found to be especially fast, as the performance was determined to be 9.4 Gflops and 12.0 Gflops on a 512 node CM-5 and CM-5E, respectively. The fluid dynamics solver was also found to scale well for large problems, however both the performance and the scaling are significantly affected by the nearest-neighbor communications, which was found to account for at least 24