The International Conference for High Performance Computing, Networking, Storage and Analysis
Optimizations of a Spectral/Finite Difference Gyrokinetic Code for Improved Strong Scaling Toward Million Cores.
Authors: Shinya Maeyama (Japan Atomic Energy Agency), Tomohiko Watanabe (National Institute for Fusion Science), Yasuhiro Idomura (Japan Atomic Energy Agency), Motoki Nakata (Japan Atomic Energy Agency), Masanori Nunami (National Institute for Fusion Science), Akihiro Ishizawa (National Institute for Fusion Science)
Best Poster Finalist
Abstract: Plasma turbulence is inherently multi-scale physics including electron and ion scales, which requires huge computations. In this work, we present optimizations of the gyrokinetic Vlasov simulation code GKV, which performs five-dimensional CFD calculation by using spectral and finite difference methods, on massively parallel platforms. First, segmented rank mapping on the three-dimensional torus network is advantageous to utilize the bi-section bandwidth and to reduce collisions of the messages, and therefore, to reduce communication costs. Second, uses of the collective communication optimized for the K computer and of the simultaneous communications on the Tofu interconnect further reduce the costs. Finally, computation-communication overlaps with pipelining methods effectively mask the communication cost. Thanks to the optimizations, GKV achieves excellent strong scaling up to ~600k cores with high parallel efficiency ~99.99994% on the K computer, and enables us to address first multi-scale plasma turbulence simulations employing the real ion-to-electron mass ratio.