Development of Performance Portable Solver Based on Kokkos Template Metaprogramming

doi:10.11871/jfdc.issn.2096-742X.2024.01.002

Abstract

Abstract:

[Objective] This paper provides a solution to programming of linear algebra problem solvers oriented to multiple heterogeneous computing architectures. [Context] The hardware architecture of supercomputers is becoming increasingly diverse. Due to incomplete software ecosystem, programming on new hardware is difficult, for example, the development cycle is too long for new hardware, repeated adaptation and development are needed for different hardware. [Methods] Using the Kokkos library to port linear algebra operators can help development of a metaprogramming framework using performance portable templates for domestic exa-scale computing environment. [Results] The Krylov sub-space algorithm are implemented as a demonstration for programming linear algebra problem solvers on a domestic heterogeneous device, resulting in at least 10x speedup compared to a 10-core hyperthreading Xeon CPU. [Conclusions] Performance portable programming is a solution for programming diverse heterogeneous computing systems.

Key words: performance portability, Kokkos, template metaprogramming, linear algebra problem solver

ZHENG Liang, LI Kunyun, ZHOU Xingbin, LI Yonghui, YU Yaojie, XIANG Yukai, HU Jian, CHAI Hua, GUO Li. Development of Performance Portable Solver Based on Kokkos Template Metaprogramming[J]. Frontiers of Data and Computing, 2024, 6(1): 12-20, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2024.01.002.

Figures/Tables 9

Fig.1

Fig.2

Fig.3

Fig.4

Fig.5

Fig.6

Fig.7

Fig.8

Table 1

References 22

[1]	MESSINA P. The exascale computing project[J]. Computing in Science & Engineering, 2017, 19(3): 63-67.
[2]	DUBEY A, MCINNES L C, THAKUR R, et al. Performance portability in the exascale computing project: exploration through a panel series[J]. Computing in Science & Engineering, 2021, 23(5): 46-54.
[3]	DEAKIN T, MCINTOSH-SMITH S, PRICE J, et al. Performance portability across diverse computer architectures[C]// 2019 IEEE/ACM International Workshop on Performance, Portability and Productivity in HPC (P3HPC). IEEE, 2019: 1-13.
[4]	EDWARDS H C, TROTT C R, SUNDERLAND D. Kokkos: Enabling manycore performance portability through polymorphic memory access patterns[J]. Journal of parallel and distributed computing, 2014, 74(12): 3202-3216. doi: 10.1016/j.jpdc.2014.07.003
[5]	EDWARDS H C, TROTT C R. Kokkos: Enabling performance portability across manycore architectures[C]// 2013 Extreme Scaling Workshop (xsw 2013). IEEE, 2013: 18-24.
[6]	BECKINGSALE D A, BURMARK J, HORNUNG R, et al. RAJA: Portable performance for large-scale scientific applications[C]// 2019 ieee/acm international workshop on performance, portability and productivity in hpc (p3hpc). IEEE, 2019: 71-81.
[7]	REGULY I Z. Evaluating the performance portability of SYCL across CPUs and GPUs on bandwidth-bound applications[J]. arXiv preprint arXiv:2309.10075, 2023.
[8]	BARATTA I, RICHARDSON C, WELLS G. Performance analysis of matrix-free conjugate gradient kernels using SYCL[C]// International Workshop on OpenCL. 2022: 1-10.
[9]	RUL S, VANDIERENDONCK H, D'HAENE J, et al. An experimental study on performance portability of OpenCL kernels[C]// 2010 Symposium on Application Accelerators in High Performance Computing (SAAHPC'10), 2010.
[10]	谭光明. 高性能计算中的性能工程问题[J]. 数值计算与计算机应用, 2022, 43(4): 343. doi: 10.12288/szjs.s2022-0842
[11]	李韦, 文渊博, 孙广中, 等. 提升高性能计算程序性能可移植性的领域特定语言[J]. 高技术通讯, 2020, 30(2): 141-149.
[12]	HUANG X, HUANG X, WANG D, et al. OpenArray v1. 0: a simple operator library for the decoupling of ocean modeling and parallel computing[J]. Geoscientific Model Development, 2019, 12(11): 4729-4749. doi: 10.5194/gmd-12-4729-2019
[13]	XU M, LI H, SONG Z, et al. An Electric-Thermal-Solid Physical Fields Coupling Calculation Based on FELAC Platform[C]// 2022 IEEE 5th International Conference on Electronics Technology (ICET). IEEE, 2022: 532-536.
[14]	ABRAHAMS D, GURTOVOY A. C++ template metaprogramming: concepts, tools, and techniques from Boost and beyond[M]. New York: Pearson Education, 2004.
[15]	TROTT C R, HAMMOND S D, MOORE S G, et al. Sustainability and Performance thorugh Kokkos: A Case Study with LAMMPS[R]. Sandia National Lab.(SNL-NM), Albuquerque, NM (United States), 2016.
[16]	MOORE S G. LAMMPS Kokkos Package--Targeting Next-Generation HPC Platforms[R]. Sandia National Lab.(SNL-NM), Albuquerque, NM (United States), 2016.
[17]	EDWARDS H C, Trott C R, Sunderland D. Kokkos Tutorial: A Trilinos package for manycore performance portability[R]. Sandia National Lab.(SNL-NM), Albuquerque, NM (United States), 2013.
[18]	MILLS R T, ADAMS M F, BALAY S, et al. Toward performance-portable PETSc for GPU-based exascale systems[J]. Parallel Computing, 2021, 108: 102831. doi: 10.1016/j.parco.2021.102831
[19]	CIESKO J, POLIAKoff D, HOLLMAN D S, et al. Towards generic parallel programming in computer science education with Kokkos[C]// 2020 IEEE/ACM Workshop on Education for High-Performance Computing (EduHPC). IEEE, 2020: 35-42.
[20]	TROTT C, BERGER-VERGIAT L, POLIAKOFF D, et al. The kokkos ecosystem: Comprehensive performance portability for high performance computing[J]. Computing in Science & Engineering, 2021, 23(5): 10-18.
[21]	LIU Y, SCHMIDT B. LightSpMV: Faster CSR-based sparse matrix-vector multiplication on CUDA-enabled GPUs[C]// 2015 IEEE 26th International Conference on Application-specific Systems, Architectures and Processors (ASAP). IEEE, 2015: 82-89.
[22]	LANGR D, TVRDIK P. Evaluation criteria for sparse matrix storage formats[J]. IEEE Transactions on parallel and distributed systems, 2015, 27(2): 428-440. doi: 10.1109/TPDS.2015.2401575

设备	国产异构加速卡	AMDVEGA20	NVIDIAV100
核心频率	1,400 Mhz	1,400 MHz	1,380 Mhz
双精度浮点性能	5.9 FLOPS	6.53 FLOPS	7.0 FLOPS