Multitasking for Local Parallelism in Applications To Chemically Reacting Supersonic Flows On Cray Y-Mp

Abstract

This article is concerned with research on parallel pro cessing for chemically reactive supersonic flows. Effi ciency of autotasking and microtasking is examined for the problems of linear algebra and two-dimensional Euler equations with the finite element method. For microtasking, satisfactory improvements in speedup (2.63 for 4 CPUs and 1.67 for 2 CPUs) are achieved, with the load balance and the amount of parallelism being the dominant factors for the speedup. Autotask ing has advantages in providing automatic paralleliza tion, but it has been shown that in the solution of finite element Euler equations based on conventional algo rithms, parallelization by means of user's intervention is required for significant speedups (3.35 for 4 CPUs and 1.79 for 2 CPUs). A much better performance may be achieved with a combination of autotasking and microtasking. These results are verified from various example problems of linear algebra and supersonic diffuser flow field calculations with hydrogen-oxygen chemical reactions.