A Parallel Computing Framework for Dynamic Power Balancing in Adaptive Mesh Refinement Applications

This chapter describes a new computing paradigm for irregular applications, in which the computational load varies dynamically and neither the nature nor location of this variation is known a priori. In such instances, load imbalances between distributed processes become a serious impediment to parallel performance, and different schemes have to be devised to balance the load. This chapter describes preliminary work, which uses “computational power balancing”. In this model, instead of balancing the load on distributed processes, the processes ask for help by recruiting other processors. This chapter illustrates the feasibility and practicality of this paradigm in an adaptive mesh application for solving non-linear dynamical systems and in the solution of large linear systems using an Additive Schwarz Preconditioned Conjugate Gradient method. The paradigm is illustrated on a SGI-Cray Origin 2000 system using MPI for distributed programming and OpenMP for shared memory programming.