Speeding Up Finite Element Wave Propagation for Large-Scale Earthquake Simulations
Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-10-109, October 2010.
Ricardo Taborda, Julio López, Haydar Karaoglu, John Urbanic*, Jacobo Bielak
Parallel Data Laboratory
Computer Science Department
Carnegie Mellon University
Pittsburgh, PA 15213
This paper describes the implementation and performance of a new approach to finite element earthquake simulations that represents a speedup factor of 3x in the total solving time employed by Hercules--the octree-based earthquake simulator developed by the Quake Group at Carnegie Mellon University. This gain derives from applying an efficient method for computing the stiffness contribution at the core of the solving algorithm for the discretized equations of motion. This efficient method is about 5 times faster than our previous conventional implementation. We evaluate the performance and scalability of the new implementation through numerical experiments with the 2008 Chino Hills earthquake under various problem sizes and resource conditions on up to 98K CPU cores, obtaining excellent results. These experiments required simulations with up to 11.6 billion mesh elements. The newly obtained efficiency reveals that other areas in Hercules, such as inter-processor communication, waiting time, and additional computing processes become more critical, and that improvements in these areas will result in significant enhancement in overall performance. This latest advance has enormous implications for saving CPU hours and catapults the potential of Hercules to target larger and more realistic problems, taking full advantage of the new generation of petascale supercomputers.
FULL TR: pdf