Parallelization of the Multilevel Fast Multipole Algorithm (MLFMA) on Heterogeneous CPU-GPU Architectures

Levent Gurel, University of Illinois at Urbana-Champaign

Usage Details

In general, our goal is to solve extremely large electromagnetics problems involving big data issues. To accomplish that, one tool that we develop is the multilevel fast multipole algorithm (MLFMA) and another tool that we use is a parallel supercomputer.

In this project, our goals are

• to port and adapt our parallel MLFMA codes to the homogeneous CPU-only Blue Waters architecture,
• to further parallelize our MLFMA codes on the heterogeneous CPU-GPU Blue Waters architecture, and
• to get ready for (and perhaps attempt, depending on the preliminary results) the solution of unprecedentedly large problems.

GPU-incorporated implementation of MLFMA for the solution of very large-scale problems will be studied for the first time, since earlier attempts reported in the literature involve relatively small problems, with at most a few millions of unknowns. We will develop and employ novel parallelization techniques specific to the heterogeneous CPU-GPU architecture of Blue Waters.