Accelerating an efficient implementation of MLFMA with hererogeneous computing

This thesis investigates possible optimization on an efficient implementation of the multilevel fast multipole algorithm (MLFMA), which is intended for solving integral equations for large problems. Though MLFMA is not inherently parallel due to its tree-like computational structure, if carefully optimized, it is suitable for parallelization as the throughput and computation power becomes higher on current GPU accelerators. By dividing problems into hierarchical multilevel groups, the MLFMA can be distributed to supercomputers like the Blue Waters, utilizing massive computing resources and balancing the workload. For solving large problems with stability and fast convergence rate, several different iterative solvers are written using PETSc (Portable, Extensible Toolkit for Scientific Computation) math library routines in the MLFMA and compared for performance. The use of GPU accelerators has also been implemented in CUDA C++ and showed great improvement on Blue Waters.U of I OnlyUndergraduate senior thesis not recommended for open acces