GPGPU性能模型研究

J4 ›› 2013, Vol. 35 ›› Issue (12): 1-7.

• 论文 • 下一篇

GPGPU性能模型研究

王锋，杜云飞，陈娟

（国防科学技术大学计算机学院，湖南长沙 410073）

收稿日期:2013-08-10 修回日期:2013-10-12 出版日期:2013-12-25 发布日期:2013-12-25
基金资助:
国家863计划资助项目（2012AA010903）；国家自然科学基金资助项目（61170049）

Research on GPGPU performance models

WANG Feng,DU Yunfei,CHEN Juan

（School of Computer Science,National University of Defense Technology,Changsha 410073,China）

Received:2013-08-10 Revised:2013-10-12 Online:2013-12-25 Published:2013-12-25

摘要/Abstract

摘要：

GPGPU的发展为并行程序带来了丰富的计算资源，但是对程序优化提出了更高的要求。程序性能模型对定位程序性能瓶颈，指导优化方法，平衡与其他设备的负载等方面起着重要作用。描述了当前性能模型的研究现状，并对其进行分类和分析。总体上性能模型分为基于统计方法的性能模型和性能解析模型，性能解析模型又分为性能度量模型、计算和访存并行性感知的模型和分部件定量分析性能模型。每种模型都给出了优缺点，并且实现了一个基于统计信息的插值性能模型，用于指导负载平衡。最后对存在的问题和未来的挑战进行了阐述。

关键词: GPGPU, GPU, 性能模型

Abstract:

The emerging and the development of the GPGPU afford the massive computation power to the parallel applications. How to use this computation power efficiently relies on the optimization of the applications. The performance models play an important roles on the targeting the performance bottleneck, guiding the optimization strategies, load balancing with other devices, etc. The stateofart of the GPGPU performance model is described, categorized, and analyzed in details. The models are divided into statisticbased curvefitting model and analysis model. The latter one can further be divided into metric model, computation and memory parallel aware model and componentbased quantitative model. The pros and cons of all the models are analyzed, and an interpolation performance model based on statistics is implemented. Finally the unsolved problems and the future challenges are presented.

Key words: GPGPU;GPU;performance model

王锋，杜云飞，陈娟. GPGPU性能模型研究[J]. J4, 2013, 35(12): 1-7.

WANG Feng,DU Yunfei,CHEN Juan. Research on GPGPU performance models [J]. J4, 2013, 35(12): 1-7.

参考文献

［1］NVIDIA. GPUs are only up to 14 times faster than CPUs says Intel, 2010［EB/OL］.［20100601］http:∥blogs.nvidia.com/ntersect/2010/06/gpusareonlyupto14timesfasterthancpussaysintel.html.
［2］http:∥www.top500.org.
［3］Luk CK, Hong S, Kim H. Qilin:Exploiting parallelism on heterogeneous multiprocessors with adaptive mapping ［C］∥Proc of the 42nd Annual IEEE/ACM International Symposium on Microarchitecture, 2009:4555.
［4］Hong S, Kim H. An integrated GPU power and performance model ［C］∥Proc of the 37th Annual International Symposium on Computer Architecture, 2010:280289.
［5］Sim J, Dasgupta A, Kim H, et al. A performance analysis framework for identifying potential benefits in GPGPU applications ［C］∥Proc of the 17th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, 2012:1122.
［6］Samadi M, Hormati A, Mehrara M, et al. Adaptive inputaware compilation for graphics engines ［C］∥Proc of the 33rd ACM SIGPLAN Conference on Programming Language Design and Implementation, 2012:1322.
［7］Peng Di,Xue Jingling.Modeldriven tile size selection for DO ACROSS loops on GPUs ［C］∥Proc of the 17th International Conference on Parallel Processing, 2011:401412.
［8］Choi J W, Singh A, Vuduc R W. Modeldriven autotuning of sparse matrixvector multiply on GPUs ［C］∥Proc of the 15th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, 2010:115126.
［9］http:∥en.wikipedia.org/wiki/Graphics_processing_unit.
［10］Owens J D,Luebke D,Govindaraju N,et al.A survey of generalpurpose computation on graphics hardware［C］∥Proc of Eurographics 2005,2005:2151.
［11］Wang Feng, Yang Canqun, Du Yunfei, et al. Optimizing linpack benchmark on GPUaccelerated petascale supercomputer ［J］. J Comput Sci Technol, 2011,26(5):854865.
［12］Yang Canqun,Wang Feng,Du Yunfei,et al.Adaptive optimization for petascale heterogeneous CPU/GPU computing ［C］∥Proc of IEEE International Conference on Cluster Computing, 2010:1928.
［13］QuintanaOrtí G, Igual F D, QuintanaOrtí E S,et al. Solving dense linear systems on platforms with multiple hardware accelerators ［C］∥Proc of the 14th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, 2009:121130.
［14］Linderman M D, Collins J D, Wang Hong, et al. Merge:A programming model for heterogeneous multicore systems ［C］∥Proc of the 13th International Conference on Architectural Support for Programming Languages and Operating Systems, 2008:287296.
［15］http:∥www.khronos.org/opencl.
［16］http:∥developer.nvidia.com/cuda/cudatoolkit.
［17］http:∥www.openaccstandard.org.
［18］Kim Jungwon, Seo Sangmin, Lee Jun, et al. SnuCL: An OpenCL framework for heterogeneous CPU/GPU clusters ［C］∥Proc of the 26th ACM International Conference on Supercomputing, 2012:341352.
［19］Ryoo S, Rodrigues C I, Stone S S, et al. Program optimization space pruning for a multithreaded GPU［C］∥Proc of the 6th Annual IEEE/ACM International Symposium on Code Generation and Optimization, 2008:195204.
［20］Hong S, Kim H. An analytical model for a GPU architecture with memorylevel and threadlevel parallelism awareness ［C］∥Proc of the 36th Annual International Symposium on Computer Architecture, 2009:152163.
［21］Yuan Liang,Zhang Yunquan,Long Guoping,et al.A GPU computational model based on latency hidden factor［J］. Journal of Software, 2010,21(zk):251262. (in Chinese)
［22］Zhang Y, Owens J D. A quantitative performance analysis model for GPU architectures ［C］∥Proc of IEEE 17th International Symposium on High Performance Computer Architecture, 2011:382393.
附中文参考文献：
［21］袁良, 张云泉, 龙国平, 等. 基于延迟隐藏因子的GPU计算模型［J］. 软件学报, 2010,21(增刊):251262.

[1]	张勇, 张曦, 万云博, 何先耀, 赵钟, 卢宇彤. 非结构有限体积CFD计算的网格重排序优化[J]. 计算机工程与科学, 2022, 44(10): 1721-1729.
[2]	朱浩, 周博洋, 卢雪山, 杜溢墨. OpenCL计算软件栈评估[J]. 计算机工程与科学, 2021, 43(12): 2105-2114.
[3]	罗磊, 陈照云, 王俪璇. 用户QoS感知的GPU集群深度学习任务动态调度[J]. 计算机工程与科学, 2021, 43(08): 1331-1340.
[4]	龚昊, 刘莹, 冯建周, 赵仁良, 冷佳旭, . 基于GPU加速的脉冲多普勒雷达信号处理[J]. 计算机工程与科学, 2021, 43(07): 1141-1149.
[5]	孙兆鹏, 周宽久. 基于PCIe的高性能FPGA-GPU-CPU异构编程架构[J]. 计算机工程与科学, 2021, 43(04): 641-651.
[6]	魏金晖, 李晨, 鲁建壮. 多GPU系统虚实地址转换架构研究[J]. 计算机工程与科学, 2021, 43(02): 228-234.
[7]	张亮, 常旭, 秦志楷, 沈立. 量子线路模拟器QuEST在多GPU平台上的性能优化[J]. 计算机工程与科学, 2021, 43(01): 17-23.
[8]	李荣春，周鑫，潘衡岳，牛新，高蕾，窦勇. 基于GPU的并行Turbo乘积码译码器[J]. 计算机工程与科学, 2020, 42(05): 761-769.
[9]	徐频捷1,2，王诲喆1,2，李策3，唐丹1，赵地1. 基于脉冲神经网络与移动GPU计算的图像分类算法研究与实现[J]. 计算机工程与科学, 2020, 42(03): 397-403.
[10]	左宪禹1,2，张哲1,5，黄祥志4,5，葛强1,2，张理涛3，臧文乾4,5. 一种适用于GPU图像处理算法的合并存储结构[J]. 计算机工程与科学, 2020, 42(02): 197-202.
[11]	王吉军，郝子宇，李宏亮. 3D-MMA:基于3D集成电路的矩阵乘加速结构[J]. 计算机工程与科学, 2019, 41(12): 2110-2118.
[12]	安婷玉1，郭宝宝2. 基于国产软硬件的OpenCL计算平台研究[J]. 计算机工程与科学, 2019, 41(11): 1919-1923.
[13]	沈立，杨耀华，王志英. 通过部分Warp重组消除GPGPU控制流的不一致性[J]. 计算机工程与科学, 2019, 41(08): 1335-1342.
[14]	张诗情，杨耀华，沈立，王志英. 通过部分页迁移实现CPU-GPU高效透明的数据通信[J]. 计算机工程与科学, 2019, 41(07): 1168-1175.
[15]	孙婷婷，黄皓，王嘉伦，翁楚良. 面向CPU-GPU异构系统的数据分析负载均衡策略[J]. 计算机工程与科学, 2019, 41(03): 417-423.

GPGPU性能模型研究

Research on GPGPU performance models

PDF

可视化

摘要/Abstract

引用本文

使用本文

参考文献

相关文章 15

编辑推荐

Metrics

本文评价