Wednesday 09 Dec 2015: Escaping the traps of strong and weak scaling with parallel-in-time integration
Daniel Ruprecht - University of Leeds
Laver 212 15:30-16:30
For problems involving the solution of time-dependent partial differential equations, computer simulations on massively parallel machines parallelise by decomposing the spatial domain. While this approach has been used with tremendous success, for the extreme number of cores in future systems it is most likely not going to be enough.
Recently, methods to additionally parallelise along the time axis have gained popularity. I will introduce one example of such a 'parallel-in-time' integration method, the 'parallel full approximation scheme in space in time' PFASST. Performance results for PFASST from massively parallel simulations will be shown and strategies to interweave PFASST with iterative spatial solvers and to incorporate fault recovery strategies will be illustrated. PFASST is based on spectral deferred corrections (SDC) and I will show some recent results how the iterative nature of SDC can be exploited to construct e.g. high order splitting schemes.