An energy-efficient parallel multigrid method for multi-core CPU platforms and HPC clusters

Abstract

For many applications, multigrid methods can be considered as the method of choice to solve or precondition linear systems of equations in the context of high performance computing. However, maintaining efficiency in the parallelization of the method for distributed memory platforms by means of a domain decomposition is not straight forward due to the different problem sizes of the grids in the multigrid hierarchy. We propose the use of an adaption strategy which adjusts the hardware activity according to the solver needs dynamically during runtime. With a focus on multi-core CPU platforms and clusters, we show that our technique can improve the parallel performance and reduce the energy consumption of the multigrid solver due to a temporary deactivation of CPU cores.