Coarse Grain Task Parallelization of Earthquake Simulator GMS Using OSCAR Compiler on Various Cc-NUMA Servers

Abstract

This paper proposes coarse grain task parallelization for a earthquake simulation program using Finite Difference Method to solve the wave equations in 3-D heterogeneous structure or the Ground Motion Simulator (GMS) on various cc-NUMA servers using IBM, Intel and Fujitsu multicore processors. The GMS has been developed by the National Research Institute for Earth Science and Disaster Prevention (NIED) in Japan. Earthquake wave propagation simulations are important numerical applications to save lives through damage predictions of residential areas by earthquakes. Parallel processing with strong scaling has been required to precisely calculate the simulations quickly. The proposed method uses the OSCAR compiler for exploiting coarse grain task parallelism efficiently to get scalable speed-ups with strong scaling. The OSCAR compiler can analyze data dependence and control dependence among coarse grain tasks, such as subroutines, loops and basic blocks. Moreover, locality optimizations considering the boundary calculations of FDM and a new static scheduler that enables more efficient task schedulings on cc-NUMA servers are presented. The performance evaluation shows 110 times speed-up using 128 cores against the sequential execution on a POWER7 based 128 cores cc-NUMA server Hitachi SR16000 VM1, 37.2 times speed-up using 64 cores against the sequential execution on a Xeon E7-8830 based 64 cores cc-NUMA server BS2000, 19.8 times speed-up using 32 cores against the sequential execution on a Xeon X7560 based 32 cores cc-NUMA server HA8000/RS440, 99.3 times speed-up using 128 cores against the sequential execution on a SPARC64 VII based 256 cores cc-NUMA server Fujitsu M9000, 9.42 times speed-up using 12 cores against the sequential execution on a POWER8 based 12 cores cc-NUMA server Power System S812L.