Performance Research of a Class of Stencil Applied in Many-Core NUMA Architecture

doi:10.11871/jfdc.issn.2096-742X.2023.06.006

Abstract

Abstract:

[Application Background] Stencil is a typical algorithm for scientific computing such as CFD (Computational Fluid Dynamics), and its memory access performance has attracted attention. The NUMA architecture is widely used in the ARM architecture represented by the Kunpeng 920 processor due to its good scalability. [Methods] Performance analysis tools and benchmark programs are used to test the performance of the Kunpeng platform's memory access and communication subsystems. The hot spot analysis and performance test are carried out for the typical stencil application software CCFD V3.0, and the Roofline model is established. [Results] The Kunpeng 920 processor relies on its many-core NUMA architecture. Its single-node floating-point performance, peak memory bandwidth, and communication latency are better than that of the Intel Xeon E5-2680v2 and another domestic processor. On a single node, the execution speed of CCFD V3.0 on the Kunpeng platform is about 2~3 times of that of the Intel platform and 1.5~2 times of that of the domestic processor. [Conclusions] The Kunpeng platform based on the ARM architecture is easy in program porting, and its NUMA architecture has advantages for memory-intensive applications such as stencil.

Key words: Stencil, Kunpeng 920, performance evaluation, CFD

GAO Lingyun, GOU Wenjin, LIU Xiazhen, YUAN Wu, ZHANG Jian, LU Zhonghua. Performance Research of a Class of Stencil Applied in Many-Core NUMA Architecture[J]. Frontiers of Data and Computing, 2023, 5(6): 58-66, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2023.06.006.

Figures/Tables 12

Fig.1

Fig.2

Table 1

Fig.3

Fig.4

Fig.5

Fig.6

Fig.7

Fig.8

Fig.9

Fig.10

Table 2

Alg.1 Schematic diagram of LU-SGS algorithm"

Algorithm 1 LUSGS
function LUSGS(nbl)
for $i c o m p = 1 → i d i m + j d i m + k d i m - 5$ do
Store Front Information
end for
for $i c o m p = 1 → i d i m + j d i m + k d i m - 5$ do
!$ omp parallel for
for $i p t s = 1 → n p t s [i c o m p]$ do
$m a t r i x ← m a t r i_l$
$r h s ← m a t r i x$
$r e s ← r h s$
end for
!$ omp parallel end for
end for
for $i c o m p = i d i m + j d i m + k d i m - 5 → 1$ do
!$ omp parallel for
for $i p t s = 1 → n p t s [i c o m p]$ do
$m a t r i x ← m a t r i_u$
$r h s ← m a t r i x$
$r e s ← r h s$
end for
!$ omp parallel end for
end for
$r e s → r e s$ * jacobian
return res
end function

Table 2

References 15

[1]	RAWAT P S, VAIDYA M, SUKUMARAN-RAJAM A, et al. Domain-Specific Optimization and Generation of High-Performance GPU Code for Stencil Computations[J]. Proceedings of the IEEE, 2018, 106(11): 1902-1920. doi: 10.1109/JPROC.2018.2862896
[2]	LI K, YUAN L, ZHANG Y, et al. Reducing Redundancy in Data Organization and Arithmetic Calculation for Stencil Computations[C]. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2021: 1-15.
[3]	YUAN L, CAO H, ZHANG Y, et al. Temporal Vectorization for Stencils[C]. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2020: 1-13.
[4]	YOSHIDA T. Fujitsu high performance CPU for the Post-K Computer[C]. Hot Chips 2018, 30(1): 1-22.
[5]	魏嘉, 张兴军, 纪泽宇, 等. 天河三号原型机分布式并行深度神经网络性能评测及调优[J]. 计算机工程与科学, 2021, 43(5): 782-791.
[6]	WIKICHIP. TaiShan v110-Microarchitectures-HiSilicon[EB/OL].[2019-05-02]. https://en.wikichip.org/wiki/hisilicon/microarchitectures/taishan_v110.
[7]	MCCALPIN, JOHN D. Memory bandwidth and machine balance in current high performance computers[J]. IEEE computer society technical committee on computer architecture (TCCA) newsletter, 1995, 2:19-25.
[8]	MVAPICH. MPI OVER INFINIBAND, Omni-Path, Ethernet/iWARP, and RoCE - Network Based Computing Laboratory, OSUMicro-Benchmarks[EB/OL]. [2022-02-03]. http://mvapich.cse.ohio-state.edu/benchmarks/.
[9]	WEAVER V. M. Self-monitoring overhead of the Linux perf_event performance counter interface[C]. IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), 2015: 102-111.
[10]	刘夏真, 马文鹏, 张鉴, 等. 基于多块结构网格大规模并行计算的负载均衡设计及实现[J]. 科研信息化技术与应用, 2013, 4(5): 18-25.
[11]	刘夏真. 并行流场软件-CCFDv3.0设计及面向国产异构平台的实现[D]. 北京: 中国科学院大学, 2021.
[12]	YAMAMOTO Y, KAI T, HOZUMI K. Numerical rebuilding of aerothermal environments and CFD analysis of post flight wind tunnel tests for hypersonic flight experiment HYFLEX[C]. AIAA Thermophysics Conference, 2001, 2899: 1-16.
[13]	龚春叶, 包为民, 汤国建, 等. 二维结构化网格CFD LU-SGS时间推进并行算法[J]. 计算机科学与探索, 2013, 7(10): 936-943.
[14]	WILLIAMS S, WATERMAN A, PATTERSON D. Roofline: an insightful visual performance model for multicore architectures[J]. Communications of the ACM, 2009, 52(4): 65-76. doi: 10.1145/1498765.1498785
[15]	BROWNE S, DONGARRA J, GARNER N, et al. A portable programming interface for performance evaluation on modern processors[J]. The international journal of high performance computing applications, 2000, 14(3): 189-204. doi: 10.1177/109434200001400303

处理器类型	Intel E5-2680 V2	Kunpeng920	国产处理器
架构	Ivy Bridge	ARMv8-A	x86_64
制程	22nm	7nm	-
主频	2.8GHz	2.6GHz	2.0GHz
核心数量	10	64	32
L3 Cache	25MB	64MB	-
内存	4x DDR3-1866	8x DDR4-2933	-
发布时间	2014	2019	-