 |
Center for Nonlinear Studies |
Parareal is a relatively new parallelization technique that focuses on the time coordinate in order to parallelize it and allows us to use more processors than conventional techniques would. Parareal is an iterative process with two stages in each iteration. In a first predictor stage, a fast coarse time propagator gives an approximate solution for all time. In a second stage, an accurate time propagator is used in order to correct the solution. The first stage is fast and computed sequentially. The second one is expensive but can be computed in parallel. The key to success in the application of the algorithm to an specific problem is to chose an adequate coarse solver. In this talk, we will introduce the Parareal algorithm and demonstrate its application to two convection-dominated problems: a 2D drift-wave simulation using the BETA code, and a 5D gyrokinetic simulation using the GENE code. Partial success was reported previously [1-3]. Here, a new and promising coarse solver based on semilagrangian time advance is proposed and tested in both applications.
[1] D. Samaddar, D.E. Newman, R. Sanchez, J. Comput. Phys. 229, 6558 (2010)
[2] J.M. Reynolds-Barredo, D.E. Newman, R. Sanchez, D. Samaddar, L.A. Berry, W.R. Elwasif, J. Comput. Phys. 231, 7851 (2012)
[3] J.M. Reynolds-Barredo, D.E. Newman, R. Sanchez, J. Comput. Phys. 255, 293 (2013)
Host: Luis Chacon, T-5