PECASE: Modeling, Analysis and Control of Subexponential Traffic Streams in Multimedia Networks

越来越多的经验证据表明，在现代多媒体网络中，流量流具有亚指数、重尾、长距离依赖和自相似的统计特征。从可变比特率（VBNS）压缩视频到各种Web和Internet相关应用，都有具有次指数特征的流量流的例子。具有这些特征的流量流与传统的语音流截然不同。这种差别不仅是数量上的，而且是质量上的。例如，当在慢时间尺度上观察语音流量时，语音流量平均并显示出很小的波动，而VBNS视频流量在多个时间尺度上显示出动态统计行为；这通常被称为自相似行为。此外，对亚指数随机序列的典型行为的大偏差以与指数流（如语音）完全不同的方式发生。在设计高效可靠的有线或无线多媒体通信网络时，理解这些新颖而有趣的统计现象是至关重要的。提高通信网络效率的最常见方法是通过共享网络资源，例如链路容量和缓冲空间。这个过程导致服务质量（QoS）的下降，这是由单个网络用户感知的。设计目标是在QoS约束下实现效率最大化，例如时延分布和损失概率。这个设计过程自然分为以下三个方面：1)统计数据分析和流量建模；2)绩效分析与控制；3)网络设计数值工具和软件工具的构建。这三个领域的研究本质上是相互关联的，任何一个领域的有效解决方案都必须以其他两个领域的进展为基础。因此，提出了一个全面、综合的研究议程。本提案的重点是对发展当前技术很重要的基本问题，例如Atm。，基于TCP/ ip和无线数据网络，以及潜在地适用于未来的网络设计。建议的研究范围从概率论到软件工程。这种影响不仅局限于多媒体通信网络，而且对于理解亚指数模型在运筹学、金融和保险理论中发挥重要作用的问题也很重要。除了提出的研究之外，还提出了一个将理论与计算相结合的课程。伟大的科学进步常常是通过理论和实验的相互作用而发生的。从这个角度来看，研究人员认为工程师在职业生涯的早期以统一的方式接触理论和实验知识是至关重要的。这使他们能够认识到理论和实验的潜力和局限性，并成功地利用它。这种方法还通过在现实世界问题上展示理论结果来加快学习过程，并教会学生如何从实验观察中提取基本信息，并将其转化为永恒的设计原则。

Increasing empirical evidence has demonstrated the presence of subexponential, heavy-tailed, long range dependent and self-similar statistical characteristics of traffic streams in modern multimedia networks. Examples of traffic streams with subexponential characteristics range from variable bit rate (VBNS) compressed video to various Web and Internet related applications. Traffic streams with these characteristics differ radically from conventional voice streams. The difference is not only quantitative, but qualitative as well. For example, when observed on a slow time scale voice traffic averages out and shows very little fluctuations while VBNS video traffic exhibits a dynamic statistical behavior on multiple time scales; this is often termed self-similar behavior. Furthermore, large deviations from the typical behavior of a subexponential random sequence occur in a completely different manner than with an exponential (e.g., voice) stream. Understanding these novel and intriguing statistical phenomena is of utmost importance when designing efficient and reliable wired, or wireless, multimedia communication networks.The most common way of increasing the efficiency of communication networks is through the sharing of network resources, e.g., link capacity and buffer space. This process leads to the degradation of quality of service (QoS) that is perceived by individual network users. The design objective is to maximize the efficiency subject to the QoS constraints, e.g., delay distributions and loss probabilities. This design process naturally breaks down into the following three areas: 1) statistical data analysis and traffic modeling; 2) performance analysis and control; 3) construction of numerical and software tools for network design. These three areas of research are inherently interconnected and an effective solution used for any one of them must be based on the advances in the other two areas. Therefore, a comprehensive and integrated research agenda is suggested.The focus of this proposal is on fundamental problems that are important for developing current technologies, e.g., Atm., TCP/IP-based and wireless data networks, as well as being potentially applicable to future network designs. The scope of the proposed research ranges from probability theory to software engineering. The impact is not only limited to multimedia communication networks, but is also important for understanding the problems in operations research, finance and insurance theory where subexponential models play an important role.In addition to the proposed research, a curriculum that integrates theory with computing is presented. Great scientific advances have often occurred through an interplay of theory and experiment. From this perspective the researcher believes it is essential that engineers are exposed early in their careersto theoretical and experimental knowledge in a unified fashion. This enables them to recognize the potential and limitations of both theory and experimentation and use it successfully to their advantage. This approachalso speeds up the learning process by demonstrating theoretical results on real world problems and teaches students how to abstract the essential information from the experimental observations into ever-lasting design principles.