GPGPU性能模型研究

Research on GPGPU performance models

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。

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.