基于Web的GPU服务平台的实现

发布时间：2018-05-29 21:31

本文选题：图形处理单元 + 远程访问GPU　；参考：《北京化工大学》2013年硕士论文

【摘要】：GPU(Graphic Processing Unit)通用计算技术是近年来出现的一种新的、很有发展潜力的并行计算技术。利用GPU通用计算技术可以完成大规模的并行计算任务。它利用GPU特殊的底层硬件架构和软件平台，可以成千上百倍提高现有并行计算算法的效率。因此它在通用计算领域和市场中占有越来越多的比重。由于GPU通用计算技术在软件和硬件方面持续进步，特别是NVIDIA(英伟达)推出CUDA(Compute Unified Device Architecture)计算平台之后，越来越多的高性能计算系统和数据中心都引入了GPU。另一方面web技术也得到了长足的发展，特别在推出web service的概念之后，位于不同机器的不同应用无须通过任何第三方就可以交换数据。但是，目前对于如何通过Web Service向其他应用程序提供GPU的计算能力的这个问题尚无标准的解决方案。本文利用Web Service技术、GPU远程访问技术(RCUDA)以及持续集成技术(continuous integration)对通过webservice向其他应用程序提供GPU的计算能力的方法和接口进行了研究。我们设计了一个可以持续构建GPU应用程序的平台，优化了webservice调度GPU应用程序的队列模型。主要进行研究工作有以下几个方面： 1、搭建rCUDA的环境，采用linux的自动化编译工具链编写能自动化编译和执行CUDA应用程序的脚本。实现可以在没有GPU的调度节点机器上进行远程自动化编译CUDA应用程序和执行CUDA程序，并比较在本地执行和远程执行的效率。 2、实现对自动化脚本的WEB封装，采用maven技术来重新定义CUDA应用的工程结构。通过maven的pom.xml文件封装自动化脚本。并将其和具有maven工程结构的GPU应用程序导入到集成服务器中，使用户可以通过定制的集成服务器来自动化编译和在线执行CUDA应用程序。 3、实现自动化编译和构建CUDA应用程序的多线程调度队列模型，使用户可以通过WEB的方式并发的调度自动化编译任务。并同时实现在线执行CUDA应用程序的多线程调度队列模型，使用户可以通过web的方式在不同的GPU上执行CUDA应用程序并获得结果。 4、实现了通过xml-rpc提供的Web Service去访问GPU应用程序的接口。这部分就是将前面的功能通过xml-rpc技术封装，供用户在他们自己的应用程序中使用GPU应用程序。
[Abstract]:GPU(Graphic Processing Unit is a new and promising parallel computing technology in recent years. Large scale parallel computing tasks can be accomplished by using GPU general computing technology. It can improve the efficiency of existing parallel computing algorithms hundreds of times by using the special underlying hardware architecture and software platform of GPU. Therefore, it occupies more and more proportion in general computing field and market. Due to the continuous progress in software and hardware of GPU general computing technology, especially after Nvidia (Nvidia) introduced CUDA(Compute Unified Device Architecture) computing platform, more and more high-performance computing systems and data centers have introduced GPU. On the other hand, web technology has made great progress, especially after the introduction of the concept of web service, different applications located in different machines can exchange data without any third party. However, there is no standard solution to the problem of how to provide GPU computing power to other applications via Web Service. In this paper, the method and interface of providing computing power of GPU to other applications through webservice are studied by using Web Service and continuous integration. We design a platform to build GPU applications continuously and optimize the queue model of webservice scheduling GPU applications. The main research work has the following aspects: 1. Build the environment of rCUDA, use linux's automatic compiling tool chain to write the script that can compile and execute CUDA application program automatically. It can compile CUDA application program and execute CUDA program automatically on the machine without GPU, and compare the efficiency between local execution and remote execution. 2. The WEB encapsulation of automatic script is realized, and the engineering structure of CUDA application is redefined by maven technology. Encapsulate the automation script through maven's pom.xml file. It and the GPU application program with maven engineering structure are imported into the integration server, so that the user can compile and execute CUDA application automatically through the custom integration server. 3. Realize automatic compilation and build multi-thread scheduling queue model of CUDA application program, so that users can schedule automatic compilation tasks concurrently through WEB mode. At the same time, the multi-thread scheduling queue model for online execution of CUDA applications is implemented, which enables users to execute CUDA applications on different GPU in the way of web and obtain results. 4, through the Web Service provided by xml-rpc to access the GPU application interface. This section encapsulates the previous functionality through xml-rpc technology for users to use GPU applications in their own applications.
【学位授予单位】：北京化工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP393.09;TP338.6

【参考文献】