76. Acceleration of the Boundary Element Library BEM4I on the Knights Corner and Knights Landing Architectures
Authors: Michal Merta (Technical University of Ostrava)Jan Zapletal (Technical University of Ostrava)
Abstract: The aim of the poster is to present acceleration of the boundary element method (BEM) by the Intel Xeon Phi technology. Since the classical BEM produces dense matrices, it is particularly suitable for acceleration on manycore architectures due to data locality and the possibility to employ vectorization and optimized dense BLAS routines. Efficient implementation of BEM is necessary to enable its application to large engineering problems. The poster provides brief overview of BEM followed by the discretization approach and efficient numerical assembly of the BEM matrices (full and sparsified). We discuss the parallelization by OpenMP in shared memory and the SIMD vectorization necessary to exploit the full potential of the Xeon and Xeon Phi architectures. An interface to the domain decomposition library Espreso enables solution of large scale problems in distributed memory. We present numerical experiments performed both on the Knights Corner coprocessor and the Knights Landing standalone processor.
Two-page extended abstract: pdf