Abstract: Communication Overlap Techniques for Improved Strong Scaling of Gyrokinetic Eulerian Code beyond 100k Cores on the K-Computer

摘要

A plasma turbulence research based on 5D gyrokinetic simulations is one of the most critical and demanding issues in fusion science. To pioneer new physics regimes both in problem sizes and in time scales, an improvement of strong scaling is essential. Overlap of computations and communications is a promising approach in improving strong scaling, but it often fails on practical applications with conventional MPI libraries. In this work, this classical issue is revisited, and resolved by communication overlap techniques, which work even on conventional MPI libraries. These techniques dramatically improve the parallel efficiency of a gyrokinetic Eularian code GT5D on the K-computer and the Helios, which are based on dedicated and commodity networks, respectively. On the K-computer, excellent strong scaling is confirmed beyond 100k cores with keeping the peak ratio of ~10% (~307 TFlops using 196,608 cores), and simulations for ITER-size fusion devices are significantly accelerated.