Fifth Parallel-in-time Integration Workshop (16w5030)
Matthew Emmett (Computer Modelling Group Ltd)
Ronald Haynes (Memorial University of Newfoundland)
Michael Minion (Lawrence Berkeley National Lab)
Martin Gander (Université de Genève)
Rolf Krause (Università della Svizzera italiana)
Scientific computing is an increasingly important tool in many areas of science and engineering. By simulating models of physical phenomena we can, for example, gain insight into processes that are difficult or impossible to measure experimentally.
The efficient use of modern high performance computing (HPC) systems has become one of the key challenges in computational science. Top HPC architectures have already attained million-way concurrency, and current trends suggest that processor counts will continue to grow rapidly. Exploiting these levels of parallelism using traditional techniques for spatial parallelism becomes problematic when, for example, the problem size per processor shrinks and communication costs begin to dominate.
For the numerical solution of time-dependent differential equations, time-parallel methods have recently been shown to provide a promising way to extend prevailing strong-scaling limits. To overcome the seemingly inherent serial dependence in the time direction, and to enable integration of multiple time-steps simultaneously, one idea of time-parallel methods is to introduce a space/time hierarchy, where integrators with different accuracies and costs are coupled in an iterative fashion. Serial dependencies are shifted to the coarsest level, allowing the computationally expensive parts on finer levels to be treated in parallel. Thus, the key to optimal parallel performance in most time-parallel methods is a well-balanced relationship between aggressive coarsening and fast convergence.
This workshop brings together scientists from the fields of parallel-in-time integration, multigrid methods, and domain decomposition to discuss similarities between their respective approaches to space/time algorithms, their applications and, ultimately, their combination. Questions arising in this context range from possible coarsening strategies to suitable inter-level transfer operators in space and time, and how other features of time-parallel methods can be exploited to further enhance their applicability to complex systems.
This workshop is the fifth "Parallel-in-time integration" workshop, and will build upon the success of the preceding four workshops:
- Università della Svizzera italiana; Lugano, Switzerland; 2011;
- Manchester University; Manchester, UK; 2013;
- Forschungszentrum Jülich; Jülich, Germany; 2014.
- Technische Universität Dresden; Dresden, Germany; 2015;
Recent advances in parallel-in-time integration, in terms of both theoretical aspects and practical advances, will be showcased by participants through a series of keynote and contributed talks.
Our intended participants include young researchers as well as domain experts in time-parallel, multigrid, and domain decomposition methods. The workshop will include several hands on sessions to develop theoretical understanding of the techniques and approaches involved in these fields and give participants a chance to further explore opportunities for collaboration.
Parallel-in-time integration methods have recently moved from proof-of-concept to production applications through the development of several software libraries that implement the MGRIT, Parareal, PFASST, and RIDC algorithms. The workshop will include hands on sessions to walk potential library users through the process of incorporating and using time-parallel methods in their codes. Library designers will also have an opportunity to exchange best-practises, design principles, pitfalls etc.