SC13 Denver, CO

The International Conference for High Performance Computing, Networking, Storage and Analysis

Solution of the Time-Dependent Acoustic-Elastic Wave Equation on a Heterogeneous, Coprocessor-Enabled Supercomputer.

Authors: Jesse Kelly (University of Texas at Austin), Hari Sundar (University of Texas at Austin), Omar Ghattas (University of Texas at Austin)

Abstract: Modern supercomputers increasingly utilize accelerators and coprocessors. A key challenge in using such heterogeneous systems has been achieving load balance such that neither the CPU nor the coprocessor is left idle. Traditional approaches have offloaded entire computations to the coprocessor, resulting in an idle CPU, or have opted for task-level parallelism requiring large data transfers between the CPU and the coprocessor. True work-parallelism has been hard as coprocessors cannot directly communicate with other CPUs (besides the host) or other coprocessors. We present a new nested partition scheme to overcome this problem. By partitioning the work assignment on a given node asymmetrically into boundary and interior work, and assigning the interior to the coprocessor, we are able to achieve excellent efficiency while ensure proper utilization of both the CPU and the coprocessor. The problem used for evaluating such partitioning is an hp-discontinuous Galerkin spectral element method for coupled acoustic-elastic wave propagation.

Poster: pdf
Two-page extended abstract: pdf

Poster Index