Copyright Hari Balakrishnan, 1998-2005, all rights reserved. Please do n ot redistributewithout permission.LECTURE 9Principles Underlying Internet QoSThis lecture discusses the main principles underlying Int ernet QoS. We don’t focusmuch on specific schemes (IntServ and DiffServ), leaving that for the sections at theend.The material for these notes is drawn partially from:1. S. Shenker [She95], Fundamental Design Issues for the Future Internet , IEEE Journalon Selected Areas in Communications, Vol. 13, No. 7, Sept ember 1995, pp. 1176-1188.2. D. Clark, S. Shenker , and L. Zhang, Suppor ting Real-Time Applications in an Inte-grated Services Packet N etwork: Architecture and Mechanisms, in Proc. SIGCOMM’92, Baltimore, MD, August 1992.! 9.1 Motivation: Why QoS?In the previous th ree lectures, we studied how the In ternet manages two network re-sources: link bandwidth and queue buffer space. A key feature of how the Internet man-ages these resources is that it relies heavily on end-system cooperation. The lecture onfair queueing showe d how network switches could isolate flows from each other usingseparate queues and non-FIFO link sched uling.Armed with link scheduling and qu eue management methods, we can think about suit-able architectures for improving on the Internet’s best-effort service mo del. That’s what QoSis all about, and research on this topic includes both architecture and specific schedulingalgorithms. This lecture will focus on the key architectural approaches, rather than onalgorithms (variants of fair queueing suffice for our discussion).QoS research was motivated by the perceived need for the network to provide morethan the best-effort service. The arguments for why the net work should provide somethingbetter than best-effort are usually made along the following lines:1. The best-effort model serves all applications, but is not optimal for many applica-tions. In particular, applications that require a minimum rate (e.g., some forms ofvideo), or delay bounds (e.g., telephony), or bounded variation in delays (e.g., live12 LECTURE 9. PRINCIPLES UNDERLYING INTERNET QOSconferencing, perhaps streaming me dia) could all benefit from being isolated fromother less-demanding applications.2. Revenue opportun ities of ISPs: If ISPs were able to treat certain customers or certainapplication flows better than others, they might make more money. This argumenthas become a critical d eterminant of whether QoS is deployed in any network or not:only schemes that provide clear economic incentives and revenue opportunities arelikely to be deployed by n etwork providers.3. Traffic overload protection: Denial-of-service attacks, flash crowds, and certainbandwidth-hungry applications (e.g. , P2P file s h aring applications today) all con-sume large amounts of bandwidth and prevent other traffic from receiving their“share”. ISPs are interested in isolating classes of traffic from each other, usuallyby setting rate controls on traffic classes and/or providing guaranteed bandwidth tocertain other classes. QoS mechanisms tur n out to help in meeting traffic manage-ment and overload protection goals.! 9.1.1 What kind of Qo S might we want?From t h e standpoint of the application provider, the following goals might make sense:1. Guaranteed bandwidth for certain application flows.2. Guaranteed delay bounds for certain application flows.3. Minimizing jitter (the variation in delay).4. Low or close-to-zero packet loss.From t h e standpoint of the network provider, the following goals might make sens e:1. Traffic classes for different customer flows, such as “Gold”, “Silver”, “Bronze”, etc.,all priced differently. For example, a provider might sell a customer conten t providercompany “Gold” service at a certain rate (and perhaps delay) for all its port 80 trafficas long as the rate of traffic does not e x ce ed some boun d specified in the service levelagreement (SLA).2. SLAs that specify minimum rate guarante es through the ISP network for custome rtraffic, relative rate shares between traffic classes, maximum rates through a net workfor traffic classes (for protection), delay priorities, etc.3. Under some conditions, the network provider might want to control who ge ts touse the network. I f the provider does this, he is said to implement admission control.Some QoS architectures fundament ally rely on admission control, but others don’t.1We will look at this issue later in this lecture.Although th e goal o f improving application performance using QoS is laudable, thatalone turns out to be p oor motivation for anyone to deploy QoS. The only chance QoS has1Recall that the circuit-switched telephone network implements admission control.SECTION 9.2. DEBATING THE FUTURE SERVICE MODEL: BEST-EFFORT V. RESERVATIONS 3of being deployed in practice is if it is strongly tied to an improved revenue stream forISPs.It took network researchers close to a decade to fully realize this point. The initialapproach to network QoS, called Integrated Services (IntServ) did not make significant prac-tical impact because it did not have a clear plan for how ISPs would make money withit. The natural question is why an ISP should deploy complex machinery for a contentprovider to make money from end clients. The lack of a good answer to this questionmeant that this architecture did not have the right deployment incentives, even thoughthe IntServ model had well-defined end-to-end service semantics for applications.In contrast, the Differentiated Services (DiffServ) architecture was motivated by ISP inter-ests, and defines per-hop behaviors that don’t really translate into anything well-defined forapplications. However, it allows ISPs to provide service differentiation between customersand their traffic classes.The [CF98] paper covers the Diff Serv topic, while the [CSZ92] paper covers the IntServtopic. Note that the methodology for implementing th ese two service models in the in-ternet today is a bit different from the descriptions in the paper. The papers provide anoverview of the key issues.Sections 9.4 and 9.5 discuss the details of DiffServ and IntServ. Before that, we discussa few issues in QoS service models.! 9.2 Debating the future service model : Best-effort v. reser-vationsThis section is a quick summary of Shenker’s paper [She95].! 9.2.1 Motivation• The In ternet today offers a single class of “best-effort” service. No assurance of whenor if packets will be delivered, no