Speaker

Dr. Daniel Ruprecht, Institute of Computational Science, University of Lugano, Switzerland

Contents

Time-parallel methods like Parareal or PFASST provide a possibility to introduce a new direction of parallelism in addition to commonly used spatial parallelization by domain decomposition. Implementations of such a space-time parallel approach mostly use MPI in space and time, distinguishing between both communication in time and space by defining corresponding MPI communicators. While straightforward to implement, this strategy requires communication of volume data by message passing. Using shared memory for the time parallelization can avoid this and reduce the memory footprint of the code. The talk will present a strategy for a hybrid implementation of Parareal with spatial parallelization, discuss issues like pipelining Parareal and the effects of different initializations of hybrid MPI and present numerical results.

Time

Wednesday, 5 December 2012, 16:00

Venue

Jülich Supercomputing Centre, Rotunda, building 16.4, room 301