The International Conference for High Performance Computing, Networking, Storage and Analysis
Extreme-Scale Space-Time Parallelism.
Authors: Daniel Ruprecht (University of Lugano), Robert Speck (Forschungzentrum Juelich), Matthew Emmett (Lawrence Berkeley National Laboratory), Matthias Bolten (Universität Wuppertal), Rolf Krause (University of Lugano)
Abstract: Time-parallel methods for the integration of time-dependent problems introduce concurrency in the temporal direction in addition to typically used parallelization in space.
A recently developed algorithm of this type is the "parallel full approximation scheme in space and time" (PFASST).
PFASST performs "sweeps" with a low-order time stepper along a hierarchy of space-time meshes, iteratively computing a collocation solution on multiple time-slices simultaneously.
For improved parallel efficiency, an FAS correction allows to use space-time coarsening strategies and thus to reduce the computational cost of sweeps on higher levels.
The poster presents novel scaling results of a combination of PFASST with a parallel multi-grid (PMG) solver in space on up to the full IBM Blue Gene/Q system JUQUEEN with 448K cores.
It is demonstrated that for large core counts the space-time parallel solver consisting of PFASST plus PMG features significantly better strong scaling capabilities than the pure space-parallel PMG code alone.