TY - GEN
T1 - Optimization of the Himeno Benchmark for SX-Aurora TSUBASA
AU - Onodera, Akito
AU - Komatsu, Kazuhiko
AU - Fujimoto, Soya
AU - Isobe, Yoko
AU - Sato, Masayuki
AU - Kobayashi, Hiroaki
N1 - Funding Information:
Acknowledgments. This research was partially supported by MEXT Next Generation High-Performance Computing Infrastructures and Applications R&D Program, entitled “R&D of A Quantum-Annealing-Assisted Next Generation HPC Infrastructure and its Applications”. The authors also would like to acknowledge HPC Solutions for providing Infiniband products.
Publisher Copyright:
© 2021, Springer Nature Switzerland AG.
PY - 2021
Y1 - 2021
N2 - This paper focuses on optimizing the Himeno benchmark for the vector computing system SX-Aurora TSUBASA and analyzes its performance in detail. The Vector Engine (VE) of SX-Aurora TSUBASA achieves a high memory bandwidth by High Bandwidth Memory (HBM2). The Himeno benchmark solves Poisson’s equation using the Jacobi iteration method. The kernel performs 19-point stencil calculations in the 3D domain, which is known as a memory-intensive kernel. This paper introduces four optimizations in a single VE or multiple VEs for the Himeno benchmark. First, for a single VE, to exploit the high bandwidth of the last-level cache (LLC) in the VE, the highly reusable array elements are stored in the LLC with the highest priority. Second, the computational domain is decomposed by considering the architecture of the VE so that this optimization can achieve a high LLC hit ratio and a long vector length. Third, to alleviate the loop overhead that tends to be large for vector computation, loop unrolling is applied to the kernel. Fourth, for multiple VEs, the optimization to improve the sustained MPI communication bandwidth is applied. The process mapping is optimized by considering different types of communication mechanisms of SX-Aurora TSUBASA. The evaluation results show that the optimizations contribute to the long vector length, the high LLC hit ratio, and the short MPI communication time of the Himeno benchmark. As a result, the performance and the power efficiency are improved due to efficient vector processing through the optimizations.
AB - This paper focuses on optimizing the Himeno benchmark for the vector computing system SX-Aurora TSUBASA and analyzes its performance in detail. The Vector Engine (VE) of SX-Aurora TSUBASA achieves a high memory bandwidth by High Bandwidth Memory (HBM2). The Himeno benchmark solves Poisson’s equation using the Jacobi iteration method. The kernel performs 19-point stencil calculations in the 3D domain, which is known as a memory-intensive kernel. This paper introduces four optimizations in a single VE or multiple VEs for the Himeno benchmark. First, for a single VE, to exploit the high bandwidth of the last-level cache (LLC) in the VE, the highly reusable array elements are stored in the LLC with the highest priority. Second, the computational domain is decomposed by considering the architecture of the VE so that this optimization can achieve a high LLC hit ratio and a long vector length. Third, to alleviate the loop overhead that tends to be large for vector computation, loop unrolling is applied to the kernel. Fourth, for multiple VEs, the optimization to improve the sustained MPI communication bandwidth is applied. The process mapping is optimized by considering different types of communication mechanisms of SX-Aurora TSUBASA. The evaluation results show that the optimizations contribute to the long vector length, the high LLC hit ratio, and the short MPI communication time of the Himeno benchmark. As a result, the performance and the power efficiency are improved due to efficient vector processing through the optimizations.
KW - Himeno benchmark
KW - Performance analysis
KW - Performance optimization
KW - Vector computing
UR - http://www.scopus.com/inward/record.url?scp=85103264688&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85103264688&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-71058-3_8
DO - 10.1007/978-3-030-71058-3_8
M3 - Conference contribution
AN - SCOPUS:85103264688
SN - 9783030710576
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 127
EP - 143
BT - Benchmarking, Measuring, and Optimizing - Third BenchCouncil International Symposium, Bench 2020, Revised Selected Papers
A2 - Wolf, Felix
A2 - Gao, Wanling
PB - Springer Science and Business Media Deutschland GmbH
T2 - 3rd BenchCouncil International Symposium on Benchmarking, Measuring, and Optimizing, Bench 2020
Y2 - 15 November 2020 through 16 November 2020
ER -