数据流技术在GPU和大数据处理中的应用  

The usage of dataflow model in GPU and big data processing

作　　者：苏华友[1] 梅松竹[1] 李荣春[1] 窦勇[1] SU Huayou;MEI Songzhu;LI Rongchun;DOU Yong(School of Computer,National University of Defense Technology,Changsha 410073,China)

机构地区：[1]国防科技大学计算机学院,湖南长沙410073

出　　处：《大数据》2020年第3期117-128,共12页Big Data Research

基　　金：国家重点研发计划基金资助项目(No.2018YFB1003400)。

摘　　要：数据流模型是一种高效的计算模型,由于其在并行性方面具有天然的优势,数据流技术在软硬件领域得到了广泛的应用。在硬件体系结构方面,数据流模型引领计算机体系结构在传统冯·诺伊曼架构下向支持更高并发的方向发展。基于超长向量处理单元的流处理和SIMT的现代GPU就广泛使用了数据流技术的思想。在编程模型方面,数据流思想在大数据编程模型领域得到了广泛应用,例如MapReduce和Spark等。从数据流模型的角度多层次分析了英伟达GPU的体系结构以及CUDA编程模型,阐述了数据流模型在GPU软硬件系统中的应用。分析了数据流思想和GPU大规模并行处理体系结构在大数据处理中的应用和发展趋势。Dataflow model is an efficient computing model.It has been widely used in software and hardware fields due to its natural advantages in parallelism.In terms of hardware architecture,the dataflow model leads the computer architecture to the direction of supporting higher concurrency from the traditional von Neumann architecture.The stream processor based on the long vector processing unit and the SIMT GPU are two instances of using dataflow technology.In terms of programming models,dataflow ideas have been widely used in the field of big data programming models,such as MapReduce and Spark.The architecture of NVIDIA GPU and CUDA programming model were analyzed from the perspective of dataflow model.The applying and trend of dataflow and GPU were analyzed in big data processing,and ideas and methods were provided for applying GPU-based systems to the field of big data processing.

关 键 词：数据流 GPU 大数据处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]