Communication Cost Function for Parallel CFD Using Variable Time Stepping Algorithms

Abstract

Recent parallel computational fluid dynamics (CFD) algorithm development demonstrates that variable time-stepping approach can significantly reduce the computation and communication time requirement. The variable time-stepping algorithm means that different sizes of time-steps can be selected independently in different subdomains (blocks) for each time-step. The new communication cost function has been tested on an IBM SP computer. The test case showed that the cost function could predict the execution time of variable time step parallel CFD code. This chapter describes a new communication cost function for parallel CFD using variable time stepping algorithms. The experiments demonstrate the accuracy of the proposed communication cost function. Workstation networks are widely used for parallel CFD. A unique problem in parallel CFD is load balancing. This chapter focuses on dynamic load balancing for parallel CFD on a heterogeneous and multi-user environment for several years. The approach is to cut the problem domain into “n” blocks and distribute the blocks among “m” processors, where m < n. Computer load is balanced by distributing blocks among computers such that the maximum elapsed execution time of all blocks is minimized.