August 14, 2014

Earthquake Simulation Scales to 8.6 Petaflops

Tiffany Trader
Intel Labs earthquake research fault

Intel Labs Vision Stategist Divya Kolar explores how supercomputing is helping society prepare for earthquakes. Current earthquake early warning systems are very limited, and basically only sound the alarm after the earthquake has begun. In a typical scenario, the affected population may have tens of seconds of lead time up to a minute or so depending on the local geography and how far they are from the epicenter. The greater the quake magnitude, the longer the warning window will be, but it’s not still long enough to launch a full evacuation strategy.

While building codes and drills are part of the solution to earthquake preparedness, the full answer lies in understanding the phenomenon of earthquake rupture and fault branching, writes Kolar. The ultimate goal will be to increase the prediction lead times for these devastating natural disasters, however simulating large earthquake events require powerful supercomputers, on the order of many petaflops or more.

Intel researchers and scientists at the Technical University of Munich and Ludwig Maximilian University of Munich are working diligently toward facilitating ever more realistic simulations. Key to the research is the earthquake simulation code, SeisSol, a complex, real-world application, with tens of thousands of lines of C/C++ and Fortran code.

Working in collaboration with Intel Labs Parallel Computing Lab, the German scientists used SuperMUC, Stampede and Tianhe-2 supercomputers in tandem with the SeisSol code to create high-resolution and high-order realistic 3D seismic simulations. By making architecture-aware optimizations, the team was able to scale code performance to an unprecedented level, harnessing half of the Tianhe-2 supercomputer (the entire portion that was allotted to the team) to reach a sustained 8.6 petaflops (double precision). The team’s performance model indicates 18-20 double-precision petaflops as the potential capability of the full Tianhe-2 machine.

“This amounts to the highest-ever sustained application-level performance for any supercomputing platform,” observes Kolar. “Equally noteworthy is the overall time-to-solution boost credited to the Intel Xeon Phi coprocessor used in Tianhe-2 system. The SuperMUC and the Tinahe-2 supercomputers have comparable size: 8192 nodes of Intel Xeon with Intel Xeon Phi on Tianhe-2 versus 9,216 nodes of Intel Xeon on SuperMUC. About 8x more peak performance of the Tianhe-2 machine improved overall time-to-solution for the 1992 Landers earthquake simulation scenario by about 2.7x!”

“Insights gathered from these simulations are of high relevance for scientific and industrial applications that help societies be best prepared for natural disasters such as earthquakes,” Kolar concludes.

The paper “Petascale High Order Dynamic Rupture Earthquake Simulations on Heterogeneous Supercomputers” is being considered for the esteemed ACM Gordon Bell Prize, which recognizes outstanding real-world HPC applications. The research will be presented at SC14.