High Performance CPU Kernels for Multiphase Compressible Flows
講演者: Christian Conti, Babak Hejazialhosseini, Diego Rossinelli and Petros Koumoutsakos (ETH Zurich, Switzerland)
主催： 学術国際情報センター・GPU コンピューティング研究会
協賛： JST CREST ポストペタスケール高性能計算に資するシステムソフトウェア技術の創出
場所： 学術国際情報センター・情報棟2F会議室 (キャンパスマップ)
The problem of Shock-Bubble Interaction (SBI) is of fundamental importance for disciplines such as medical research where its understanding is required for the extracorporal shock wave lithotripsy, a non-invasive technique used for the treatment of renal stones.
To study the problem we rely on high order numerical methods to solve the compressible Euler equations. However, due to the high resolutions required to capture all the relevant scales in simulations of SBI, high throughput software is needed.
In this work we present high performance kernels for multiphase compressible flows that aim at exploiting as much as possible the computing capabilities of modern multicore CPU platforms. We show how the organization of the solver in a software stack helps the optimization process, discuss the optimizations applied on each layer and perform an a-priori performance analysis in order to quantify and evaluate the measured performance.
We then present a simulation obtained with the current software and measurements performed on the various layers of the software stack. We conclude with an overview with the current and future plans for the framework.