Skip to main content
Seminar | Mathematics and Computer Science Division

On Time Integration for Strong Scalability

LANS Informal Seminar

Abstract: Modern computer architecture is characterized by increasing computing capacity (via vectorization and intranode parallelism) and modest improvements in network and DRAM bandwidth while latency has remained nearly constant. As a result, the prospects for PDE-based simulation to reduce time to solution are poor when using current algorithms. We consider a collection of related numerical methods and implementation techniques for shortening the critical path of accurate and stable transient simulation as a way to improve strong scalability and reduce memory bandwidth demands. This approach unifies some parallel in time” methods in terms of diagonal Runge-Kutta schemes, achieve different trade-offs in the multi-objective Pareto front of {accuracy, time, cost}, and confronts numerical stability challenges. We consider several implementation strategies and include a preliminary evaluation of the methods in the context of stiff and non-stiff PDEs. This seminar will be streamed.