On the scalability of parallel computations on a network of workstations

Domain decomposition (DD) is confirmed to be a useful parallel technique for parallel simulation. Mesh partition strategy together with a master/slave construction represents a natural way of doing parallel distributed computations. In a recent study using an unstructured finite volume implementation with Parallel Virtual Machine (PVM), Dou and Phan-Thien [1] found that there is a limit to the speed-up when the problem size increases to a certain value. In this paper, we will discuss the parallel strategies for DD technique, the load balancing, and the communication issue among the processors, and explain the mechanism of the limit, and search for the way to enhance the speed up and the scalability.