Breaking Global Communication -- Multi-level Parallelism for High-dimensional Problems

Dirk Pflüger, Mario Heene, University of Stuttgart,
Christoph Kowitz, Technische Universität München

We consider turbulence simulations for fusion reactors as a prominent example for high-dimensional problems. Additional to space and time, velocities have to be jointly discretized in the underlying PDEs. While high-dimensional problems have inherent need for tomorrow's supercomputers, a hierarchical splitting -- the sparse grid combination technique -- allows us to tackle several HPC issues at once. A hierarchical splitting breaks the need for global communication and synchronization, and partial failures can be compensated without checkpoint-restart and with often only little loss in accuracy. We show novel algorithms and communication schemes that make use of this hierarchical approach for high-dimensional problems in plasma physics and beyond.

Workshop: Exploiting Different Levels of Parallelism for Exascale Computing

as part of the ACM International Conference on Supercomputing - ICS 2014

Date

June 10, 2014

Venue

Bavarian Academy of Sciences, Munich

Contact

Miriam Mehl, miriam.mehl@ipvs.uni-stuttgart.de
Dirk Pflüger, dirk.pflueger@ipvs.uni-stuttgart.de