An investigation of load balancing strategies for CFD applications on parallel computers

Publisher Summary This chapter employs an Euler and Navier-Stokes code, PARC2D, to demonstrate the basic concepts of dynamic load balancing. This code solves unsteady flow equations using conservation variables and provides different order Runge-Kutta time-stepping schemes. The chapter explains the procedure for parallel implementation of the explicit time-integration algorithm. Numerical experiments were chosen to demonstrate some strategies in dynamic load balancing for managing computer systems and algorithms. The chapter describes load balancing for variations in the system loading. The test problem was run on four processors over a period of approximately twelve hours. Communication and computation costs for each process were recorded and a load balancing was performed after approximately every thirty minutes. The results of this computation for a controlled environment are summarized. The same problem is illustrated for an uncontrolled environment. Further, load balancing for heterogeneous algorithms in parallel computing is described. This type of load balancing strategy optimizes the algorithm on a parallel system, and dynamically load balances the problem as the algorithm adapts to the solution.