【作者】 徐嬴颖；

【导师】 陈常嘉； 赵永祥；

【作者基本信息】 北京交通大学 ， 通信与信息系统， 2014， 博士

【摘要】 文件分发系统是互联网重要的信息传播平台。云计算的出现和云技术的迅猛发展,促成了云下载系统的出现并迅速成为关注的焦点。云下载系统可以提供预约式文件分发服务,用户提出下载请求之后可以离线甚至关机,而不需要一直保持在线,由此节省了大量用户时间；云下载系统可以根据用户需求,按需租用云平台提供的计算、存储和带宽等资源来获取、缓存和向用户返回预约的文件,保证用户文件可用性和取回文件的速率。预约式服务与云平台的结合,赋予了云下载系统以下新特点：增加了系统与用户的交互形式；系统可以统计不同用户的文件预约请求复用文件,减少系统资源开销；系统如何租用和管理云资源,是提高文件分发效率和降低服务成本的关键。但是,现有云下载系统的资源分配策略缺乏对预约式下载服务的深度考虑,会导致存储资源浪费、用户间服务不公平、资源开销大等问题。本文针对这些问题,以云下载系统新特点作为视角,建立了相关理论模型,并设计了优化的资源分配策略,主要工作和创新点如下：(1)针对预约式服务交互过程,本文建立了系统交互模型,从理论上刻画了采用预约式服务的云下载系统运行机制,并提出了系统响应策略(何时通知用户、何时开始获取文件)。首先,分析单个用户下载过程,建立了交互模型量化交互过程中的各时间元素及其关系,并由模型理论分析,推导出用户时间开销和系统存储时间开销。其次,建立了最小化这两个时间开销的多目标优化问题,并分别求解得到最优化用户体验和最小化系统存储开销的系统响应策略。最后,由多目标优化问题的求解得到用户时间开销与系统存储时间开销之间是折中的关系,在此基础上提出了在满足用户服务质量需求的条件下最小化系统存储开销的系统响应策略。(2)针对驱动用户与系统合作的激励问题,本文设计了适用于多用户场景的用户合作激励机制,通过对同一个文件的多个预约请求的汇聚和连续服务实现文件的高效复用。首先,建立了多用户请求云下载的多用户模型,获得不同文件复用情况下的系统服务成本。其次,在用户无私和用户自私两种情况下,分析用户与系统的收益。并针对这两种用户提出了根据用户等待时间和报价提供差分服务的合作机制。最后,通过仿真和理论分析证明了这些机制有效降低了系统获取文件和缓存文件的开销,保证了用户服务的公平性。(3)针对已有的云缓存分配算法没有考虑预约式服务特点的问题,本文提出了存储容量模型,定量地给出了系统需要的云缓存量(即云存储需求下限)。首先,通过对云下载系统中云缓存设计面临问题的深入分析,揭示了用户删除行为与文件有效期是影响系统云缓存占用量的关键因素。其次,通过大规模数据分析和挖掘,发现了云缓存的存储特性(如用户请求数与缓存文件数的关系等)和用户删除文件的规律。最后,建立了存储容量模型。该模型公式化地表示了请求数、文件有效期与系统云缓存量的关系,为云缓存容量规划和文件有效期设计提供了理论依据。与基于真实数据的仿真值进行比较,模型预估的云缓存需求量误差在10%以内。(4)针对云下载系统中存储资源开销和计算资源开销之间的关系,本文在云存储需求下限的基础上建立了系统资源开销模型,提出了云缓存内容管理算法。对大规模商业云下载系统的实际运营数据进行分析,得到系统资源开销模型中的关键参数。由该模型求解获得的云缓存租用量,在满足系统的需求的条件下使云存储和计算资源的开销总和最小。然后,对大规模真实系统中的用户行为数据进.行分析,得到了用户取回行为的特征。在此基础上,提出了一种云缓存内容管理算法(F-LRU)。最后,通过大规模实际数据驱动的仿真试验,结果显示了F-LRU算法的命中率和比特命中率都高于LRU、SIZE算法。更多还原

【Abstract】 File distribution system is an important platform for efficient information dissemination on the Internet. With the rapid development of cloud-computing, cloud download system has emerged and quickly gains focus. The reservation-based file distribution service provided by cloud download system allows users to be offline or even shutdown, instead of keeping online after sending out their requests. Therefore, it could save the users a lot of time. According to users’ requests, the cloud download system can rent computing, storage, and bandwidth resource in order to ensure the availability and also the retrieving speed of requested files.The cloud download system combining reservation service and cloud platform, has many novel features such as new interaction patterns between users and system, reusing files according to the statistics of reservation requests, and renting and managing cloud resources for improving efficiency and reducing service costs. However, existing resource allocation strategies in cloud download system have several severe problems, such as waste of storage resources, unfair service, and large overhead in resource provisioning.To address the above issues, this dissertation builds theoretical models and proposes resource assignment strategies by considering the novel features of the cloud download system. The main work and contributions in this dissertation are summarized as follows:(1)Considering interaction between user and system, this dissertation builds an interaction model, which theoretically characterizes the operation mechanism of reservation services in the cloud download system, and propose system response strategy, which decides the time when to notify the user and when to start to get the file. Firstly, based on the analysis of a single user downloading process, the theoretical interaction model is built to quantify all time elements of interaction and their relations. And the user time cost and the system storage time cost can be obtained from the model. Secondly, a multi-objective optimization problem is established. By solving it, response strategies are found to optimize the user experience or minimize system storage cost. Finally, based on the tradeoff between the user time cost and the system storage time cost, we proposed a response strategy to minimum required system storage cost with meeting the quality requirement of services to the needs of the user.(2)To deal with the incentive problem of driving the user and the system cooperation, we design cooperation mechanism to achieve efficient reuse of file in multi-user scenarios. Firstly, the multi-user model is built to achieve service cost when system reuses the file differently. Secondly, the utilities of user and system are analyzed in two cases that altruistic or selfish user. In both cases, we respectively propose mechanisms, in which the system provides difference services according to waiting time and quotation of user. Finally, by simulations and theoretical analysis, it is proved that the mechanism effectively reduces the cost of acquiring and caching file, and ensures the fairness for user.(3)To deal with the issue that the existing cache allocation algorithms did not consider characteristics of reservation service, this dissertation builds the storage capacity model to quantify required cloud cache size (i.e., lower limit of cloud storage requirement). Firstly, through meticulous investigation and statistical analysis of the challenge on designing cloud cache, we firstly put forward that the critical factors affecting cloud cache occupation are the deleting behavior of user and the validity of files. Secondly, through the massive data analysis and mining, we found the cloud cache storage features (such as the relation between the user requests’number and the cached file number, etc.) and the law of user deleting file. Finally, the cloud cache size demand model is established. The model establishes the relation among requests’ number, the validity of files and cloud cache size. It provides a theoretical basis of designing cloud cache size and validity of files. Compared with the simulation based on real-trace, results show that the relative error of storage capacity model is less than10%.(4) Considering the cloud storage costs affecting the cloud computing cost, we build model to minimize the system resource cost, and propose the cloud cache management algorithm. The key parameters of the model are studied by a real large-scale commercial cloud download system. Based on the lower limit of cloud storage requirement, the model is designed to rent a cloud cache to minimize the sum of cloud storage and computing resources cost. On the basis of users’retrieving behavior characteristics found through the massive user behavior data analysis in the real-trace, we propose the cloud cache management algorithm, i.e. F-LRU. Finally, the results of simulations driven by massive real-trace prove that both the hit rate and bit-hit rate of F-LRU are higher than the LRU and SIZE algorithm.更多还原