Multilevel Parallelization of the Space Weather Modeling Framework

Gabor Toth, University of Michigan

Usage Details

The project aims at optimizing and improving multilevel parallelization of the computationally most expensive components of the Space Weather Modeling Framework: the BATS-R-US magnetohydrodynamic model and the AMPS particle-in-cell model. Both codes are massively parallel using the MPI library to distribute data and computation over thousands of computer nodes. Our goal is to combine the MPI parallelization with multithreaded OpenMP parallelization over the dozens of cores on each of the computational nodes. This hybrid approach reduces memory use relative to a pure MPI implementation and allows performing 10 to 20 times larger simulations. In addition, we will explore the use of GPUs to speed up the computationally most expensive parts of the models by a factor of 5 to 20. Combining these parallelization strategies will allow doing unprecedented simulations that go from the global scale of the magnetosphere to the small scales of kinetic plasma processes.