Special IssueDynamic QoS Mapping Control for Streaming Video inRelative Service Differentiation Networks JITAE SHIN,JIN-GYEONG KIM,JONGWON KIM AND C.-C. JAY KUOIntegrated Media Systems Center and Department of Electrical Engineering-SystemsUniversity of Southern California, Los Angeles, California 90089-2564fjitaeshi,jingyeok,jongwon,[email protected]. A dynamic quality of service (QoS) mapping control scheme, which includes feedforward and feedbackQoS control, is proposed for the differentiated services (DiffServ) networks in this work. To achieve reliable and con-sistent end-to-end video streaming with relative service differentiation, the proposed solution consists of two parts: (1)relative priority-based indexing and categorization of streaming video contentat the sending end-system and (2) dynamicand aggregate QoS mapping control with packet, session, and class-based granularity level for categorized packets at theedge of the DiffServ domain, called the video gateway (VG), based upon the load variation of the network. In particular,we focus on dynamic solutions to handle QoS demand variations of continuous media applications (e.g. varying priori-ties from aggregated/categorizedpackets) and QoS supply variations of the DiffServ network (e.g. varying loss/delay dueto fluctuating network loads). Thus, with the proposed dynamic QoS mapping control, video streaming with enhancedquality is demonstrated under a pricing model.1INTRODUCTIONInternet applications have very diverse requirements onthe network service, thus making the current best-effort(BE) Internet model less than sufficient. The emergingcontinuous media (CM) applications demand more strin-gent QoS requirements than traditional TCP-based appli-cations. Under the BE model, maintaining the end-to-endCM quality is extremely challenging due to two reasons.First, the CM stream is inherentlya variable bit rate (VBR)data stream. Second, the Internet is an unpredictable time-varying channel. Emerging differentiated services (Diff-Serv or DS) [1, 2, 3] schemes have been extensively studiedto achieve IP-QoS in recent years. It provided more con-sistent QoS services to meet CM application’s needs thanthe current best-effort Internet service in a simple and scal-able manner. On-going research efforts in DiffServ can bedivided into two classes: absolute differentiation[4, 5] andrelative differentiation [6, 7]. An absolute service differ-entiation scheme attempts to guarantee QoS for a set ofaggregated flows regardless of background traffics. Withabsolute service differentiation, per-flow QoS is typicallyachieved through admission control. A relative service dif-ferentiation scheme tries to maintain a quality gap among This research was funded in part by the Integrated Media SystemsCenter, a National Science Foundation Engineering Research Center, un-der the Cooperative Agreement No. EEC-9529152.DS levels (e.g. behaviour aggregates [1]) in the environ-ment, where traffic loads of different behavior aggregateschange dynamically. As the Internet evolves towards theDiffServ model and as networked CM applications becomemore network-aware and adaptive, relative service differ-entiation will be more attractive due to its simplicity andflexibility.The DiffServ architecture can bring benefits to bothend-users and ISP by providing better service quality forCM applications at the willingness of users to pay morefor better services. Thus, for DiffServ-aware applicationswhich include video streaming as a special example, thearchitecture design should consider interests of both end-users and ISP. That is, an end-user should get a betterprice/performance tradeoff for his/her DiffServ-aware ap-plication while ISP benefits in providingdifferent chargingand service policies to maximize end-users’ satisfaction. Inorder to perform negotiation, we should measure the QoSdemand of CM applications and the QoS supply of Diff-Serv networks in terms of predefined granularity. With apredefined granularity, service differentiation can be de-manded by tagging (or marking) source units differentlyat the end-system. These units will be treated differentlyaccording to their tags. Furthermore, these tags can be ad-justed (i.e. re-marking) dynamically according to trafficconditions and handled differently according to new tags.Generally speaking, there are three differentiationgran-Vol. 12, No. 3, May-June 2001 217J. Shin, J.G. Kim, J.W. Kim, C.-C.J. Kuoularities. At the application side, a end-user may want todemand differentiation per session or per packet dependingon application requirements. The session-based schemematches per-flow QoS control within the access network.This DiffServ granularityhas also been promoted under thename of user-based allocation [8] and user-share differen-tiation [9]. Differentiation can however be carried out atthe packet level to enable intra-flow differentiation, i.e. thepacket-based scheme, which will be promoted in this pa-per. For packet-based marking, individual packets are pri-oritized with respect to each other for classification, queu-ing, rate-limiting, and so on. For example, sessions of CMapplications usually require further granularity for a bettermatch with the different priority of each packet in termsof quality as employed in video streaming [10, 11]. In ad-dition, there is class-based service differentiation associ-ated with the DiffServ network. That is, beyond the accessand the boundary networks, the DiffServ domain handlesonly aggregated flows of the class-based granularity, i.e.flows with the same DS levels. The class-based DiffServscheme demands some adaptive packet forwarding mecha-nism [6, 7, 12] for consistent and proportional service dif-ferentiation.Under the relative service differentiationparadigm, thiswork presents a solution in which streaming video senders,receivers, and a special boundary node called the videogateway (VG), located at the border of the DiffServ do-main, controlQoS of streaming videos interacting with oneanother under a cost constraint. In this method, a videoapplication at the source assigns chunks of its content (of-ten in the unit of packets) by certain indices according totheir impacts to end-to-end QoS in terms of loss and delay.Since these indices reflect the desired service preference ofone portion with respect to other portions, we call it therelative priority index (RPI), which can be further dividedinto two indices, i.e. the relative loss index (RLI) and
View Full Document