Regulation, Scheduling, and Flow Control for Quality of Service Support

拟议的项目将研究在有效支持不同服务质量(Qos)方面的一些问题。由于网络资源的稀缺和新应用的出现，当前的互联网必须从尽力而为的服务演变为支持不同服务质量的服务。IETF首先在综合服务的背景下探讨了这个问题，最近又在区分服务的背景下探讨了这一问题。综合服务工作的关键问题之一是在网络核心中缺乏每微流调度的可伸缩性，而区分服务的基本原则一直是将每微流操作移动到网络的边缘，并且仅处理网络内部的宏流。然而，仍然需要一些关键业务元素，例如调节器(用于整形、监管和标记边缘的微流以及不同子网络边界的宏流)和调度器(跨越不同的宏流和服务类别)来提供不同的服务质量。更高的多路复用收益也是通过从统计数据而不是在最糟糕的情况下进行考察来寻求的。此外，还提出了在微流级别使用端到端流量控制和管制/标记与在宏流级别使用丢弃(RIO)/调度相结合的方案，以提供保证的服务质量与使用附加可用带宽的能力的组合。此建议的目标是设计这些关键流量控制元素-调节器、调度器和流量控制机制-以便不仅提供所需的不同QoS，而且以最大化网络效用的方式实现这一点。为了实现这一目标，该项目将依赖于研究人员及其合作者最近开发的建模和分析框架。该框架为调节器和调度器提供了统一的数学模型(基于流量包络/服务曲线/服务进程的概念)，并且提供了一种简单的方法来分析具有(或不具有窗口/速率流控制)的此类元素的分叉-连接网络。特别是，研究人员先前的工作已经展示了如何使用简单的组合规则将整个网络的影响降低到具有(端到端)服务曲线的等价单个元素的影响，该(端到端)服务曲线是根据各个元素的服务曲线给出的。此外，这样的网络的最坏情况和概率端到端性能然后很容易根据该端到端服务曲线获得。由于该框架是可分析的，因此可以将分析结果“倒置”，以推导出设计高效交通控制元件的规则。更有趣的是，该框架允许将单个交通控制元素的复杂服务曲线表示为简单服务曲线的网络。然后可以利用这种表示来合成复杂但高效的调节器和调度器。因此，该建模框架为分析、设计和综合区分服务网络的流量控制元素提供了一种系统的方法。特别是，拟议的项目将导致以下结果：它将导致合成更有效的调节器和调度器。它将在区分服务网络的背景下开发其他有保证的服务(除了租用线路仿真)。它应该为概率服务产生更好的接纳控制方案。对于保证服务和概率服务，它还将通过定价将网络资源有效地分配到不同流量的服务曲线中。在具有某些最低要求的自适应服务的情况下，这将导致更好地理解简档内和简档外流量之间的相互作用，以及如何利用这种相互作用来增强现有的自适应机制，例如TCP。

The proposed project will investigate a number of problems in the efficient support of different qualities of service (QoS). Due to the scarcity of network resources and the emergence of new applications, it is considered crucial that the current internet evolve from a best effort service to one that supports different QoS. This problem was explored in the IETF first in the context of integrated services and more recently in the context of differentiated services. One of the key problems with the work in integrated services was the lack of scalability of per-microflow scheduling in the core of the network, and the basic tenet of differentiated services has been to move per-microflow operations to the edge of the network and deal with only macroflows in the interior of the network. Nevertheless, some of the key traffic elements such as regulators (to shape, police, and mark microflows at the edge as well as macroflows at the boundary of different subnetworks) and schedulers (across different macroflows and service classes) are still needed to provide the different QoS. Higher multiplexing gains are also sought by looking at flows statistically rather than in the worst case. Moreover, schemes that use a combination of end-to-end flow control and policing/marking at the microflow level and dropping (RIO)/scheduling at the macroflow level have also been proposed to provide a combination of assured QoS along with the ability to use additional available bandwidth. The objective of this proposal is to engineer these key traffic control elements - regulators, schedulers, and flow control mechanisms - so as to not only deliver the different QoS desired, but to do so in a manner that maximizes the network utility. In order to accomplish this objective, the project will rely on a modeling and analysis framework that has recently been developed by the researcher and his collaborators. This framework provides a unified mathematical model (based on the notion of traffic envelopes/ service curves / service processes) for regulators and schedulers as well as an easy way to analyze fork-join networks of such elements with (or without window/rate flow control). In particular, the researcher's prior work has shown how the impact of the entire network may be reduced to that of an equivalent single element with an (end-to-end) service curve given in terms of the service curves of the individual elements using an easy composition rule. Furthermore, the worst-case and probabilistic end-to-end performance for such networks is then easily obtained in terms of this end-to-end service curve. Since the framework is analyzable, one can "invert" the results of the analysis to derive rules for the design of efficient traffic control elements. More interestingly, the framework allows for the representation of a sophisticated service curve for a single traffic control element, as a network of simple service curves. This representation can then be exploited to synthesize complex but highly efficient regulators and schedulers. This modeling framework therefore provides a systematic way to analyze, design, and synthesize traffic control elements for differentiated services networks. In particular, the proposed project will lead to the following: It will result in the synthesis of more efficient regulators and schedulers. It will develop other guaranteed services (besides leased line emulation) in the context of differentiated services networks. It shall produce better admission control schemes for probabilistic service. For both guaranteed and probabilistic service, it will also lead to efficient allocation of network resources into service curves for different flows via pricing. In the case of adaptive service with certain minimum requirements, it will lead to a better understanding of the interplay between the in-profile and out-of-profile traffic, and how this interplay may be exploited to enhance existing adaptive mechanisms such as TCP.