Optimization of the Himeno Benchmark for SX-Aurora TSUBASA
Optimization of the Himeno Benchmark for SX-Aurora TSUBASA
复制标题
SX-Aurora TSUBASA 姬野基准优化
DOI：
10.1007/978-3-030-71058-3_8
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Kobayashi Hiroaki
中科院分区：
文献类型：
--
作者：
Onodera Akito;Komatsu Kazuhiko;Fujimoto Soya;Isobe Yoko;Sato Masayuki;Kobayashi Hiroaki
本文重点对矢量计算系统SX-Aurora TSUBASA的Himeno基准进行了优化，并对其性能进行了详细分析。SX-Aurora TSUBASA的Vector Engine （VE）通过high bandwidth memory （HBM2）实现高内存带宽。Himeno基准使用Jacobi迭代法求解泊松方程。内核在3D域中执行19点模板计算，这被称为内存密集型内核。本文介绍了针对Himeno基准测试在单个VE或多个VE中的四种优化。首先，对于单个VE来说，为了充分利用VE的最后一级缓存（last-level cache， LLC）的高带宽，重用度高的数组元素存储在优先级最高的LLC中。其次，考虑VE的体系结构，对计算域进行分解，使优化能够实现高LLC命中率和长向量长度；第三，为了减轻向量计算往往较大的循环开销，将循环展开应用于内核。第四，针对多ve，采用优化方法提高持续MPI通信带宽。通过考虑SX-Aurora TSUBASA不同类型的通信机制，优化了过程映射。评价结果表明，优化后的Himeno基准具有较长的矢量长度、较高的LLC命中率和较短的MPI通信时间。因此，由于通过优化的有效向量处理，性能和功率效率得到了提高。
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.
NVIDIA GPU 与 NEC SX-Aurora TSUBASA 矢量处理器中使用的 SIMD 处理功能之间的关系分析
DOI：
10.1007/978-3-030-25636-4_10
发表时间：
2019
影响因子：
5.3
作者：
I. Afanasyev;V. Voevodin;V. Voevodin;Kazuhiko Komatsu;Hiroaki Kobayashi
通讯作者：
Hiroaki Kobayashi
可重定向优化编译器中的积极循环展开
DOI：
--
发表时间：
1996
期刊：
International Conference on Compiler Construction
影响因子：
--
作者：
J. Davidson;S. Jinturkar
通讯作者：
S. Jinturkar
GPU加速计算——从炒作到主流，矢量计算的重生
DOI：
--
发表时间：
2009
期刊：
影响因子：
--
作者：
Satoshi Matsuoka;Takayuki Aoki;Toshio Endo;Akira Nukada;Toshihiro Kato;Atushi Hasegawa
通讯作者：
Atushi Hasegawa