异构系统并行计算软件性能测评分析与实证研究

doi:10.11871/jfdc.issn.2096-742X.2024.03.013

数据与计算发展前沿 ›› 2024, Vol. 6 ›› Issue (3): 116-126.

CSTR: 32002.14.jfdc.CN10-1649/TP.2024.03.013

doi: 10.11871/jfdc.issn.2096-742X.2024.03.013

异构系统并行计算软件性能测评分析与实证研究

顾蓓蓓^1,²(),邱霁岩^1,²,迟学斌^1,^*()

1.中国科学院计算机网络信息中心，北京 100083
2.中国科学院大学，北京 101408

收稿日期:2023-11-01 出版日期:2024-06-20 发布日期:2024-06-21
通讯作者: *迟学斌（E-mail：chi@sccas.cn）
作者简介:顾蓓蓓，中国科学院计算机网络信息中心，博士研究生，高级工程师，主要研究方向为并行计算，高性能计算软件性能测评分析。
本文中主要负责方法研究、实验设计、论文写作与格式校正等。
GU Beibei, Computer Network Information Center, Chinese Academy of Sciences, doctoral student, her main research directions include parallel computing, performance evaluation and analysis of high performance computing software applications.
In this paper, she is responsible for method research, paper writing, and format correction.
E-mail: gbb@cnic.cn|迟学斌，中国科学院计算机网络信息中心，博士生导师，研究员，博士，主要研究方向为并行计算。
本文中主要负责论文框架和实验指导。
CHI Xuebin, Computer Network Information Center, Chinese Academy of Sciences, doctoral supervisor, researcher, PhD, his main research directions include high performance computing, parallel computing and software.
In this paper, he is responsible for paper framework and experimental guidance.
E-mail: chi@sccas.cn
基金资助:
国家自然科学基金(62372428)

Analysis and Empirical Research on Performance Evaluation of Parallel Computing Software in Heterogeneous Systems

GU Beibei^1,²(),QIU Jiyan^1,²,CHI Xuebin^1,^*()

1. Computer Network Information Center, Chinese Academy of Science, Beijing, 100083, China
2. University of Chinese Academy of Sciences, Beijing, 101408, China

Received:2023-11-01 Online:2024-06-20 Published:2024-06-21

摘要/Abstract

摘要：

【目的】“并行计算软件性能测评”一直是超算领域重要的研究方向。在异构系统上对计算软件实际性能进行真实测评和分析，可以有效促进对异构系统计算软件生态的良性发展。【方法】本文首先通过调研文献对国内外并行计算软件性能测评方法进行研究分析，归纳总结出业界对并行计算软件性能测评的研究划分的3个重要阶段；通过并行计算矩阵乘积Cannon算法对软件的真实性能进行实证分析，并对运行时间和效率等重要指标进行多维度的实验分析。【结果】在同一节点下，不是使用越多的加速卡越能降低该并行程序的运行时间；不同规模矩阵在不使用加速卡和使用单个加速卡两种情况下，程序的并行效率均没有因为节点的增多而发生明显的变化。【结论】在异构系统计算软件中只关注并行效率不能如实反映该软件性能的真实水平。除节点间并行效率因素外，节点内部加速也成为反映并行计算软件真实水平的一项重要测评指标。

关键词: 异构系统, 并行软件, 性能测评, 并行效率

Abstract:

[Objective] The Performance evaluation of parallel computing software has always been an important research direction in the field of supercomputing. Real evaluation and analysis of the actual performance of computing software on heterogeneous systems can effectively promote the healthy development of the computing software ecosystem in heterogeneous systems. [Methods] This article first conducts research and analysis on parallel computing software performance evaluation methods, both domestically and internationally, through literature research, and summarizes three important stages of industry research on parallel computing software performance evaluation. Empirical analysis is conducted on the real performance of the software through the Cannon algorithm for parallel computation of matrix product, and multi-dimensional experimental analysis is conducted on important indicators such as execution time and efficiency. [Results] Under the same node, the execution time is not always reduced with more accelerator cards used. In the cases of using single or none accelerator card, the parallel efficiency of programs dealing with different matrix scales does not change significantly with the increase of nodes. [Conclusions] In heterogeneous system, the computing software focusing only on parallel efficiency cannot truly reflect the actual level of software performance. In addition to the efficiency factor of inter-node parallelism, intra-node acceleration has also become an important evaluation indicator reflecting the true level of parallel computing software.

Key words: heterogeneous systems, parallel software, performance evaluation, parallel efficiency

顾蓓蓓, 邱霁岩, 迟学斌. 异构系统并行计算软件性能测评分析与实证研究[J]. 数据与计算发展前沿, 2024, 6(3): 116-126.

GU Beibei, QIU Jiyan, CHI Xuebin. Analysis and Empirical Research on Performance Evaluation of Parallel Computing Software in Heterogeneous Systems[J]. Frontiers of Data and Computing, 2024, 6(3): 116-126, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2024.03.013.

图/表 10

图1

表1

表2

表3

表4

表5

图2

表6

表7

图3

参考文献 33

[1]	迟学斌, 王彦棡, 王珏, 等. 并行计算与实现技术[M]. 北京: 科学出版社, 2015: 7.
[2]	MARCO Z, BROND L, STEVE T, et al. Performance Analysis Using the MIPS R10000 Performance Counters[C]. Proceedings of the 1996 ACM/IEEE conference on Supercomputing (CDROM), 1996.
[3]	BROWNE S, DONGARRA J, GARNER N, et al. A Scalable Cross-Platform Infrastructure for Application Performance Tuning Using Hardware Counters[C]. ACM/IEEE SC 2000 Conference, 2000.
[4]	RUDOLPH B, ZIEGLER H. PCL-the Performance Counter Library: A Common Interface to Access Hardware Performance Counters on Microprocessors[Z]. Germany: Research Centre Juelich GmbH, 1993.
[5]	LUC S. System and Kernel Thread Performance Monitor API Reference Guide[R]. BM RS/6000 Division, 1999.
[6]	GERNDT M, KEREKU E. Periscope: Advanced Techniques for Performance Analysis[C]. Parallel Computing: Current & Future Issues of High-End Computing, Proceedings of the International Conference ParCo 2005, 2005.
[7]	FAHRINGER T, GERNDT M. Specification of Performance Problems in MPI-Programs with ASL[C]. International Conference on Parallel Processing (ICPP’00), 2000.
[8]	JOSEPH E, WILLARD C. A new HPC technical computing benchmark: the IDC balanced rating[C]. IDC Bulletin W, 2000.
[9]	BKDG. A Programming Based Performance Counter Statistical Method for Modern High end Microprocessors[EB/OL]. [2010-04-12]. http://support.amd.com/us/Processor_TechDocs.
[10]	TALLENT N, MELLOR J, ADHIANTO L, et al. “HPCToolkit: performance tools for scientific computing.”[C]. Journal of Physics: Conference Series, 2008.
[11]	Open SpeedShop, Performance measurement tools..[EB/OL]. [2010-04-12]. http://www.openspeedshop.org/wp/.
[12]	PAPI, Performance measurement tools. [EB/OL]. [2010-04-12]. http://icl.cs.utk.edu/papi/.
[13]	SHENDE S, MALONY A. “The Tau Parallel Performance System.”[J]. International Journal of High Performance Computing Applications, 2006, 20(2): 287-311.
[14]	MARTIN B, KIM B D, JEFF D, et al. PerfExpert: An Easy-to-Use Performance Diagnosis Tool for HPC Applications[C]. 2010 ACM/IEEE International Conference for High Performance Computing, Networking, Storage and Analysis, 2010.
[15]	GERNDT M, MOHR B, LARSSON J. Evaluating OpenMP Performance Analysis Tools with the APART Test Suite[C]. Fifth European Workshop on OpenMP (EWOMP’03), 2003.
[16]	GEIMER M, WOLF F, WYLIE B J, et al. The Scalasca Performance Toolset Architecture. Concurrency and Computation[J]. Practice and Experience, 2010, 22(6), 702-719.
[17]	KNUPFER A, BRUNST H, DOLESCHAL J, et al. The Vampir Performance Analysis Tool Set. In: Tools for High Performance Computing[J]. Springer, Berlin, 2008, 139-155.
[18]	SHENDE S S, MALONY A D. The TAU Parallel Performance System[J]. International Journal of High Performance Computing Applications. 2006, 20(2), 287-311.
[19]	DIETER M, SCOTT B. Score-P-A Unified Performance Measurement System for Petascale Applications[C]. Competence in High Performance Computing, 2010.
[20]	DONGARRA J, GANNON D, FOX G, et al. The Impact of Multicore on Computational Science Software[J]. ResearchGate, 2007, 3: 3-10.
[21]	JEFF D, MARTIN B. Evaluation and Optimization of Multicore Performance Bottlenecks in Supercomputing Applications[C]. IEEE International Symposium on Performance Analysis of Systems and Software, 2011.
[22]	CHEN D, SINGH D. Fractal video compression in opencl: An evaluation of cpus, gpus, and fpgas as acceleration platforms[C]. Asia and South Pacific Design Automation Conference (ASP-DAC), 2013.
[23]	BAKHODA A, YUAN G. Analyzing cuda workloads using a detailed gpu simulator[C]. IEEE International Symposium on Performance Analysis ofSystems and Software (ISPASS), 2009.
[24]	BOR K. K, SU B Y. clspmv: A cross-platform opencl spmv framework on gpus[C]. ACM international conference on Supercomputing (ICS), 2012.
[25]	JOHN C, SETH K, et al. Performance analysis framework for high-level language applications in reconfigurable computing[C]. ACMTrans. Reconfigurable Technology, 2010.
[26]	SILVA B, BRAEKEN A. Performance modeling for fpgas: Extending the roofline model with high-level synthesis tools[C]. International Journal ofReconfigurable Computing (IJRC), 2013.
[27]	ALTERA. Implementing fpga design with the opencl standard[R]. Altera Whitepaper, 2011.
[28]	SUBIN B, AJEESH S. Automated Performance Modeling of HPC Applications Using Machine Learning[C]. IEEE Transactions on Computers, 2020.
[29]	OZAN T, EMRE A. Diagnosing Performance Variations in HPC Applications Using Machine Learning[C]. Lecture Notes in Computer Science, 2017.
[30]	JONATHAN R, JACK D. Timemory: Modular Performance Analysis for HPC[C]. Lecture Notes in Computer Science, 2020.
[31]	ZHOU K, MARK W K. Tools for top-down performance analysis of GPU-accelerated applications[C]. Proceedings of the 34^th ACM International Conference on Supercomputing, 2020.
[32]	LARISSA S, MARCIN C. Performance-Detective: Automatic Deduction of Cheap and Accurate Performance Models[C]. International Conference on Supercomputing 2022, 2022.
[33]	FELIPE T, ALVARO W. Scalable performance analysis method for SPMD applications[J]. The Journal of Supercomputing, 2022 (78): 19346-19371.

软件/工具	支持硬件	支持语言	是否需要插装	是否需要重编译	支持并行库	依赖库
IPM	CPU、GPU		否	必须	MPI、openMP	papi
PAPI	是一个接口、具体实现各异		可选	可选		GCC等编译器
Timemory	CPU、GPU	C/C++、Python、CUDA、HIP	必须	必须	MPI、openMP	GCC等编译器
PerfExpert	CPU	C/C++	否	否	MPI、openMP	PAPI、HPCToolkit
ScoreP+Vampir	CPU、GPU	C/C++、Fortran CUDA	否	必须	MPI、SHMEM、openMP、Pthread、OpenACC、OpenCL	GCC等编译器
Extrae+Paraver	CPU、GPU	C/C++、Python CUDA	无	无	MPI、 openMP、 Pthread、	Boost、dyninst、libunwind、papi、wxwidgets

gpu-per-node	运行时间（ms）
gpu-per-node	1node	4 node	9 node	16 node
0	207,787	61,900	52,268	47,465
1	404	1,477	2,410	3,832
2	402	1,206	2,685	6,632
3	404	1,148	2,906	4,320
4	405	1,187	4,412	5,933

gpu-per-node	运行时间（ms）
gpu-per-node	1node	4 node	9 node	16 node
0	491,452	140,683	119,318	107,123
1	795	2,446	3,522	6,368
2	796	2,163	3,815	8,461
3	796	1,992	4,190	8,638
4	798	2,044	4,980	6,393

gpu-per-node	运行时间（ms）
gpu-per-node	1node	4 node	9 node	16 node
0	959,978	272,543	228,397	203,791
1	1,375	4,158	5,754	8,996
2	1,374	3,400	5,248	10,842
3	1,377	3,121	4,651	7,422
4	1,379	3,103	5,046	8,278

gpu-per-node	运行时间（ms）
gpu-per-node	1node	4 node	9 node	16 node
0	1,653,833	451,597	387,890	344,775
1	2,124	6,074	7,677	12,617
2	2,123	4,959	6,919	14,441
3	2,129	4,564	7,734	9,615
4	2,124	4,531	6,617	11,206

异构系统并行计算软件性能测评分析与实证研究

Analysis and Empirical Research on Performance Evaluation of Parallel Computing Software in Heterogeneous Systems

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 10

参考文献 33

相关文章 2

编辑推荐

Metrics

本文评价

矩阵大小	并行效率
矩阵大小	4 node	9 node	16 node
9000×9000×9000	83.92%	44.17%	27.36%
12000×12000×12000	87.33%	45.76%	28.67%
15000×15000×15000	88.06%	46.70%	29.44%
18000×18000×18000	91.55%	47.37%	29.98%

[1]	李肯立,阳王东,陈岑,陈建国,丁岩. 面向人工智能和大数据的高效能计算[J]. 数据与计算发展前沿, 2020, 2(1): 27-37.
[2]	赵海涛, 孙家昶, 黎雷生, 杨文浩, 赵慧, 李会元. 适合一类复杂异构超算系统的HPL并行计算模型研究[J]. 数据与计算发展前沿, 2020, 2(1): 85-92.

矩阵大小	并行效率
矩阵大小	4 node	9 node	16 node
9000×9000×9000	6.84%	1.86%	0.66%
12000×12000×12000	8.13%	2.51%	0.78%
15000×15000×15000	8.27%	2.66%	0.96%
18000×18000×18000	8.74%	3.07%	1.05%