AMR Multi-GPU Accelerated Tsunami Simulation

Tsunamis are natural disasters that represent a real and dangerous threat specially to countries with coasts along the Pacific Ocean. At the light of the tragic events of the 2011 Earthquake and Tsunami in Japan the importance of predicting this phenomenon has gained great relevance. In order to simulate a Tsunami the Shallow Water Equations (SWE) are used, these equations although reliable can be very costly in computational terms. In order to obtain an accurate and fast result we introduce GPGPU to solve these equations, this new technology allows us to program GPUs to obtain dramatic boosts in performance. We use the CIP-CSL2 Semi-lagrangian and the Method of Characteristics to solve numerically the SWE. In order to push the envelop on performance we develop a Multi-GPU simulation on Tsubame 2.0 and to handle large scale domains (like the Pacific Ocean) we introduce an Adaptive Mesh Refinement(AMR) technique. In this way we can save memory usage from between 20 to 40% less than if not AMR was used and moreover we can obtain high performance and good scalability from the Multi-node GPU simulation, 313GFLOPS for single GPU and currently about 1.2TFLOPS for Multi-GPU with 128 cards.