The rapidly increasing number of cores in high-performance computing systems causes a multitude of challenges for developers of numerical methods. New parallel algorithms are required to unlock future growth in computing power for applications and energy efficiency and algorithm-based fault tolerance are becoming increasingly important. So far, most approaches to parallelise the numerical solution of partial differential equations focussed on spatial solvers, leaving time as a bottleneck. Recently, however, time stepping methods that offer some degree of concurrency, so-called parallel-in-time integration methods, have started to receive more attention.
I will introduce two different numerical algorithms, Parareal (by Lions et al., 2001) and PFASST (by Emmett and Minion, 2012), that allow to exploit concurrency along the time dimension in parallel computer simulations solving partial differential equations. Performance results for both methods on different architectures and for different equations will be presented. The PFASST algorithm is based on merging ideas from Parareal, spectral deferred corrections (SDC, an iterative approach to derive high-order time stepping methods by Dutt et al. 2000) and nonlinear multi-grid. Performance results for PFASST on close to half a million cores will illustrate the potential of the approach. Algorithmic modifications like IPFASST will be introduced that can further reduce solution times. Also, recent results showing how parallel-in-time integration can provide algorithm-based tolerance against hardware faults will be shown.
- Computational Mathematics and Applications Seminar