Parallel Performance of a Zonal Navier-Stokes Code on a Missile Flowfield

摘要

While in theory parallel computing can produce very high levels of performance, it remains clear that for several reasons the track record for parallel computing has been inconsistent. Some of these reasons are: (1) parallel systems are difficult to program, (2) performance gains are failing to live up to expectations, (3) many of the current generation of RISC-based parallel computers have delivered low levels of performance. This paper discusses the highly innovative techniques that were used to overcome all of these obstacles. A time-marching, Navier-Stokes code, successfully used over a decade for projectile aerodynamics, was chosen as a test case and optimized to run on modern RISC-based parallel computers. As part of the validation process, the parallelized version of the code has been used to compute the three-dimensional (3-D) turbulent supersonic flow over a generic, ogive-cylinder, missile configuration. These results were compared to those obtained with the original version of the code on a Cray C-90. Considerable performance gain was achieved by the optimization of the serial code on a single processor. Parallelization of the optimized serial code, which uses loop-level parallelism, led to additional gains in performance. Recent runs on a 128-processor Origin 2000 have produced speedups in the range of 10-26 over that achieved when using a single processor on a Cray C-90. The parallelized code was then used to compute the turbulent flow over various projectile configurations at transonic and supersonic speeds.