Live Webcast 15th Annual Charm++ Workshop

We investigate the use of the processor virtualization technique in parallelizing the multigrid algorithm on high performance computers. By doing processor virtualization, we can achieve adaptive process overlapping, better cache performance, and dynamic load balance control. We use a neighbor based virtual processor to physical processor mapping strategy and dynamically changing the number of virtual processors with the multigrid levels to maximum the performance of the multigrid solver. A V-cycle multigrid solver, is accomplished based on Charm++, a message driven parallel language developed by Parallel Programming Lab at University of Illinois at Urbana-Champaign. Numerical experiments for solving convection-diffusion type equations on structured grids are reported to show the benefits we get from processor virtualization.