A Study on QoS Control in High-Speed IP Multiservice Networks

流聚合IP业务具有多时间尺度行为。在短时间尺度上，它表现出随机性，而在长时间尺度上，它具有长程相关性。这一事实从工程的角度来看是非常重要的注意，分组丢失概率和延迟的尾部分布在排队模型中只取决于在一定的有限时间尺度上的随机特性，因此在适当的条件下，IP网络中的分组丢失概率可以基于业务在短时间尺度上的随机特性来确定。与长距离相关业务相比，它更容易处理具有随机特性的业务，并且研究者提出的分析框架可以用于定性评估QoS性能。在这项研究中，我们研究了本地随机性的出现在IP流量的机制，并获得了一种方法来确定的时间尺度范围内，本地随机性占主导地位的性能。作为这一成就的一部分，我们获得了IEICE的最佳论文奖。我们还研究了多马尔可夫到达流模型的分析方法。注意，马尔可夫到达流可以以任意精度近似地表示任何平稳点过程。特别是，我们建立了一种方法来分析队列与时间相关的服务速度。此外，我们推导出了封闭形式的解决方案的平均缓冲区内容的离散时间队列与多个马尔可夫到达流，并揭示了如何流量参数的平均缓冲区内容的影响。此外，我们还分析了M/G/1型马氏链的几何渐近性和次指数渐近性，并给出了计算其瞬态解的算法。这些研究成果为定量评价变长分组多业务IP网络性能奠定了基础。

It is known that flow-aggregated IP traffic has multiple time-scale behaviors. In short time scales, it exhibits random characteristics, while it has long-range dependence in long time-scales. This fact is very important from an engineering view point Note that the packet loss probability and the tail distribution of delay in queueing models depend only on stochastic characteristics in a certain finite time-scale, and therefore under an appropriate condition, the packet loss probability in IP networks can be determined based on random characteristics of traffic in short time-scales. Compared with long-range dependent traffic, it is easy to handle traffic with random characteristics, and the analytical framework proposed by the investigator can be utilized for qualitative evaluation of QoS performance. In this study, we investigated the mechanism of appearance of local randomness in IP traffic and obtained a method to identify the range of time scales in which local randomness dominates the performance. For a part of this achievement, we received the Best Paper Award from the IEICE. We also studied analytical approaches to queueing models with multiple Markovian arrival streams. Note that the Markovian arrival stream can represent any stationary point process approximately with an arbitrary accuracy. In particular, we establish an approach to analyzing queues with time-dependent service speed. Further we derived the closed-form solution of the mean buffer contents in discrete-time queues with multiple Markovian arrival streams, and revealed how traffic parameters have an impact on the mean buffer contents. Moreover, we analyzed the geometric and subexponential asymptotics in Markov chains of M/G/1 type and proposed an algorithm to compute their transient solutions. With all those achievements, we established the foundations of quantitative evaluation of multiservice IP networks with variable-length packets.