The quickly rising number of cores in modern high-performance computing systems plus the fact that individual processor frequencies are no longer increasing basically means that acceleration of numerical codes requires additional levels of parallelization. For time-dependent PDEs, methods that provide concurrency in the temporal direction are a promising approach that can introduce additional parallelism already on the level of the employed numerical algorithms. The talk will introduce two methods, Parareal and PFASST, and discuss their performance in large-scale parallel simulations as well as implementation aspects. Further, the stability of Parareal for convection-dominated problems is analyzed and the "Krylov-subspace-enhanced'' version of Parareal is presented.

Lehrprobe

Dem wissenschaftlichen Vortrag wird eine Lehrprobe zum Thema "Numerik III: Prinzipielle Funktionsweise von Mehrgitterverfahren für elliptische partielle Differentialgleichungen" von ca. 20 min vorangehen.

Hierzu sind auch insbesondere die Studierenden des Fachbereichs eingeladen.