【作者】 尤国华；

【导师】 赵英；

【作者基本信息】 北京化工大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 互联网的广泛普及，特别是万维网的迅猛发展，使基于Web的远程控制技术成为远程控制领域新的发展方向和研究热点。在基于Web的远程控制系统中，Web服务器是关键部件之一，直接影响着远程控制系统能否快速、准确、可靠地完成控制任务。而且，随着光纤、新一代交换机等宽带技术在互联网领域的广泛应用，网速越来越快。与之相对，Web服务器逐渐成为基于Web的远程控制系统的性能瓶颈。Web服务器性能的提升主要依靠提高硬件性能和改进Web服务器软件的请求调度算法。为了提高Web服务器的硬件性能，人们开始把多核处理器应用于Web服务器。多核处理器在Web服务器中的应用改变了传统Web服务器软件的运行环境，也产生了新的请求调度问题。因此，本课题提出了适用于多核Web服务器的动态请求调度算法，并对多核Web服务器的建模、多核Web服务器的“乒乓”效应等问题进行了研究。所做主要工作如下：（1）建立了多核Web服务器的排队网络模型。在对Web服务器的研究过程中，为了分析Web服务器的性能，已经出现了许多使用排队网络理论或Petri网理论对Web服务器建立的数学解析模型。但是这些已建立的Web服务器的性能模型大多是针对单核处理器的。因此，本课题在详细分析Web服务器动态请求处理过程的基础上，依据排队网络理论，建立了多核Web服务器的排队网络模型。该排队网络模型使用泊松过程请求流作为输入流，但也对自相似请求流的情况进行了讨论。（2）研究了多核处理器的“乒乓”效应问题。多核处理器在一个芯片上集成多个处理器核心的同时，也改变了处理器的缓存结构。多核处理器往往有多级缓存，这样就产生了缓存一致性的问题。而缓存的一致性问题往往会引起处理器核心之间的“乒乓”效应，这会严重影响多核处理器的性能。本课题在对现有操作系统的线程调度算法和典型的片上多核处理器的缓存结构进行研究的基础之上，尝试使用操作系统提供的进程/线程硬亲和性的方法消除多核处理器的“乒乓”效应问题。为了验证此方法，进行了仿真实验，实验结果证明了此方法的有效性。（3）提出了多核Web服务器中的动态请求调度算法。在对Web服务器动态请求处理过程进行深入研究之后，为了消除多核处理器中存在的“乒乓”效应问题，在充分考虑了动态请求服务时间的分布特点的基础上，本课题提出了针对多核Web服务器的动态请求调度算法。新的动态请求调度算法先对动态请求进行分类，然后进入不同的动态请求队列，而且将为同一类动态请求提供服务的服务线程分配至同一个处理器核心以避免“乒乓”效应。此外，为了保持各处理器核心之间的负载均衡，使用遗传算法规划出最佳的服务线程分配方案。（4）进行了多核Web服务器中动态请求调度算法的仿真实验。在新的动态请求调度算法的基础上，本课题开发了仿真Web服务器，泊松过程和自相似动态请求流的生成器，并使用这些工具进行了仿真实验。实验测量出了新算法的响应时间分布、平均响应时间、丢包率等关键的性能参数，并且同常用的短作业优先和先到先服务等请求调度策略的性能进行了对比。更多还原

【Abstract】 With the rapidly development of Internet and WWW, the web-basedremote control technology become a new orientation of development andhot research topic of the remote control field. Web server is the keycomponent for web-based remote control system, and decides whetherremote control system could accomplish a task quickly, accurately, andreliably. Furthermore, with the widely application of fiber and newgeneration switch in the Internet filed, the speed of network is more andmore quick. On the contrary, web server becomes the performancebottleneck of the web-based remote control system. The performance ofweb server relies on the performance of hardware and request schedulingalgorithm in web server software. In order to improve the performance ofweb server hardware, multi-core CPU is applied to web server. Theapplication of multi-core CPU to web server changes the runningenvironment of traditional web server software, and leads to new requestscheduling problems. So this paper proposed a new dynamic requestscheduling algorithm for multi-core web server, and studied the modelingof multi-core web server and ping-pong effect of multi-core web server. The research items are as follows:(1) Constructing the queueing network model for multi-core webserver. In order to analyze the performance of web server, there have beenmany performance models for web server based on queueing networktheory and Petri theory. But these performance models of web server arejust for the single-core web server. To evaluate the performance ofmulti-core web server, the thesis constructed the queueing network modelfor multi-core web server on the basis of the analysis of requestscheduling process and queueing network theory. The model uses Poissonprocess as input arrival process. At the same time, the model discusses theself-similar arrival process as well.(2) Studying the ping-pong effect of multi-core CPU. As it integratedmulti processing cores into one chip, multi-core CPU changed thestructure of caches too. There are usually multi-level caches in multi-coreCPU, which would lead to the problem of cache coherence. The problemof cache coherence is liable to cause the ping-pong effect betweenprocessor cores, which will greatly reduce the performance of multi-coreCPU. According to the research of threads scheduling policies of currentO/S and the cache structure of multi-core CPU, this thesis tried toeliminate the ping-pong effect by means of threads/processes hard affinitymethods provided by O/S. For validating the method, we do a simulationexperiment, and the experiment results proved that it is an effective method that eliminates the ping-pong effect by means of the affinity ofthreads/processes.(3) Proposing a dynamic request scheduling algorithm for multi-coreweb server. After making a thorough study in the process of handlingdynamic requests, in order to eliminate the ping-pong effect in multi-coreCPU, a new dynamic request scheduling algorithm is proposed accordingthe service time distribution of dynamic requests. The new algorithmclassifies the dynamic requests, and then the dynamic requests enter intothe different dynamic request queues. The threads that served for thesame dynamic request queue are assigned to the same processor core.Moreover, to keep the load balance between processor cores, the bestthreads assignment policy is calculated by means of Genetic Algorithm.(4) Conducting the simulation experiment based on the new dynamicrequest scheduling algorithm in multi-core web server. The simulationweb sever was developed on the basis of the new dynamic requestscheduling algorithm in multi-core web server. Furthermore, wedeveloped the sending module of Poisson arrival process and self-similararrival process. With the help of these tools, we conducted a series ofsimulation experiments. We measured the key indices of performance,such as response time distribution, mean response time, dropped rate andso on. At the same time, we do the simulation experiments based onshortest-job-first and first-come-first-served policies respectively, and compared their performance to the new dynamic request schedulingalgorithm.更多还原