摘要
在很多新兴应用领域、如传感器网络,实时监控系统等,产生的数据流是不断变化的、连续到达的、数据值可能不确定、且必须被快速处理。其中有些操作,如数据流的实时窗口连接运算,非常消耗时间,这对数据流处理系统的性能提出了严峻的挑战。目前,大多数算法采用软件优化来提高处理速度,但其性能提高有限。利用GPU(图形处理器)的高并行度、多线程、高带宽的并行处理能力,设计了一种软硬件结合的方法来加速处理数据流的窗口连接操作。在CUDA(统一计算架构)下,由CPU控制将内存中的数据传输至GPU存储器中,然后利用多线程进行并行处理。实验验证了提出的方法可以大幅度提高多数据流窗口连接的处理速度,可达到纯软件处理的50倍左右。
Data in some emerging applications, such as sensor networks, real-time monitoring systems, etc. , are always time- varying, uncertain, and continuously arriving. These data need to be quickly processed. However, some important operations are costly, such as the real-time window joins over data streams. The requirement is a high challenge for a data streams pro- cessing system. Currently, most of the algorithms use software optimization to accelerate the processing speed, yet the results are unsatisfactory. GPU （graphic processing unit） has a high processing performance as it uses multi-thread and high-band- width to process data in parallel. This paper presented a co-processing method with hardware and software to accelerate the window join operation over data streams. In CUDA （ compute unified device architecture） architecture, CPU （ central proces- sing unit） transfers data to memory of GPU to be processed in parallel. Experiment results show that this method can greatly improve the processing speed with about 50 times faster than software implementation.
出处
《计算机应用研究》
CSCD
北大核心
2014年第5期1428-1432,共5页
Application Research of Computers
基金
浙江省自然科学基金资助项目(LY13F020040)
浙江省"信息与通信工程"重中之重学科开发基金
宁波市自然科学基金资助项目(2012A610065)
关键词
图形处理器
统一计算架构
不确定数据流
窗口连接操作
graphic processing unit （GPU）
compute unified device architecture （CUDA）
uncertain data streams
windowjoins processing
