Advances in Modern Performance Analysis: A New Approach

未来的高速网络预计将支持具有不同业务流特征和服务质量(Qos)要求的数据、语音、图像和视频等多种业务。开发一种能够有效地支持这种服务质量要求的广域分组网络的设计、控制和管理框架是一个根本性的挑战。传统的网络通信量分析和服务质量保证的框架是确定性的或概率的。在确定性方法中，使用了最坏情况下的交通信封。这很可能会导致保守的资源分配策略和网络资源的低效使用。在标准概率框架中，流量通常被建模为Poisson过程的变体，如MMPP、MAP等。这些模型适用于传统的电信流量分析，但只能捕获数据中简单且有限的关联结构，研究人员建议显著扩展分析方法来解决现代计算机和通信网络中出现的复杂电信流量问题。他们的方法植根于经典的投票定理，处理一大类到达过程，包括标准的马尔可夫过程，一般的周期过程，由TIME序列(自回归和移动平均)描述的过程，甚至是由等间隔非重叠离散区间中的到达数目的联合分布描述的一般过程。特别注意推导算法解，即使在系统利用率很高的情况下也是有效的。研究人员提出的方法是统一的，它避免了根发现，并且不像矩阵分析方法或最近引入的状态空间谱分解方法，它不涉及潜在的昂贵的迭代。此外，它是无变换的，可以为进一步的分析研究提供一种简单的解形式，例如尾部行为或渐近分析。研究人员可以使用一般到达过程来进行端到端性能分析，一旦离开网络节点的过程被适当地近似或概率地结合。由于它们不限于传统的马尔可夫模型，这种方法很有可能是可行的。研究人员建议实现我们的通用算法，并将它们作为科学包提供给研究社区。

The future high-speed networks are expected to support a variety of services such as data,voice, image, and video with diverse traffic flow characteristics and Quality-of-Service (QoS)requirements. Developing a framework for the design, control and management of widearea packet networks which can effciently support such QoS requirements is a fundamentalchallenge.Conventional frameworks for the teletraffic analysis of networks and QoS guarantees areeither deterministic or probabilistic. In the deterministic approach, the worst-case trafficenvelopes are used. This most likely will lead to conservative resource allocation policiesand inefficient usage of the network resources. In the standard probabilistic frameworks,the traffic is usually modeled as a variant of Poisson process, e.g., MMPP, MAP, etc. Suchmodels are amenable to traditional teletraffic analysis but only capture simple and limitedcorrelation structures in the data.The researchers propose to significantly expand the analytical methodology in solving complex teletraffic problems arising in modern computer and communications networks. Their approachwhich is rooted in classical ballot theorem handles a large class of arrival processes includingthe standard Markovian processes, general periodic processes, processes described bytime-series (auto-regressive and moving average), or even general processes described by thejoint distribution of number of arrivals in equally-spaced non-overlapping discrete intervals.Special attention will be paid in deriving algorithmic solutions which are numerically robustand efficient even when the system utilization is high.The researchers proposed approach is unifying, it avoids root-findings, and unlike the matrix analytical approach or recently introduced state-space spectral decomposition method, it does notinvolve potentially expensive iterations. Further, it is transform-free and may provide asimple form of solution amenable for further analytical studies such as tail behavior orasymptotic analysis.The researchers general arrivals processes could be used for the end-to-end performance analysis once the departure process from a network node is appropriately approximated or probabilisticallybounded. Since they are not limited to traditional Markovian models, such an approachappears to be viable.The researchers propose to implement our general algorithms and make them available as a scientific package to the research community.