1 Computer Communications - CSCI 551 Copyright © William C. ChengCS551Integrated andDifferentiated ServicesBill Chenghttp://merlot.usc.edu/cs551-f12resource reservations (Internet: RSVP)Integrated servicesguaranteed or probabilistic bandwidth/delaygood match for real-time traffic (e.g., VOIP)Pros:prefect for VPNs (ISPs can sell "virtual pipes")too much state for backbone routersCons:difficult policy issues between AS’s?make the most use out of your bandwidth⇒ not widely deployed2What’s Next: Integrated Services Computer Communications - CSCI 551 Copyright © William C. Chengassumes some overprovisioningbest-effort and preferred (better-than-best-effort)Differentiated servicesvery simple service modelor in and out (best-effort and less-than-best-effort)easy to implement and fast (no per-flow state)Pros:ISPs can charge extra for preferredno guaranteesCons:3What’s Next: Differentiated Services Computer Communications - CSCI 551 Copyright © William C. ChengCS551Fundamental DesignIssues[Shenker95a]Bill Chenghttp://merlot.usc.edu/cs551-f124 Computer Communications - CSCI 551 Copyright © William C. ChengReservations, admission control, etc, orDo we need to extend the Internet service model (currentlybest effort)? 5Key Ideas Computer Communications - CSCI 551 Copyright © William C. ChengOverprovision and keep best effortAsks fundamental questionsSimple model, very high level viewHelps guide the thinking for a very hard questionHow do we even study this question?U(j) maps the network’s performance to the user’s level ofhappinessDoes the network make users happy?6Model: Utility and Efficacy Computer Communications - CSCI 551 Copyright © William C. ChengDefine U(j) be the utility delivered to the jth userFor example, higher bandwidth delivered to a videoapplication (up to a point) makes the user happierSimilarly, lower delay delivered to application makesuser happier... the sum of all U(j)s (the efficacy, denoted by V)Goal of network is to maximize... and increase efficacy that waymakes better use of available bandwidthIn a best-effort network, can increase bandwidth toincrease efficacy7More Bandwidth or New Service Model? Computer Communications - CSCI 551 Copyright © William C. ChengOr, for the same bandwidth, introduce newservices matched to application needsShenker: always better to add new servicesKey question: what’s the relative cost of addingbandwidth and adding new servicesbut cost of adding new services hard to estimateNo. A single network can always use leftover bandwidthto increase efficacyDo separate networks for different applications providehigher efficacy? 8Other Considerations Computer Communications - CSCI 551 Copyright © William C. ChengOtherwise, none of these arguments holdsNote: increasing efficacy does not mean increasingeveryone’s utility Service models must map application requirementsNeed to embed application knowledge inside the network(BAD!)Should applications explicitly request service, or shouldthe network determine service to deliver?9Implicit vs. Explicit Service Request Computer Communications - CSCI 551 Copyright © William C. ChengImplicit doable if number of services is small and wellknown and stable (e.g., port number)Applications must know what they want!Explicit supports larger variety of services but incentivesneeded so all do not request highest serviceMajor coordination effort (imagine changing IP orEthernet..)Stable service model needed so apps know what to requestPricing, accounting and billing: these are hardOverload: a network is overloaded if by removing a flowwould increase V even though there are fewer flows10Admission Control? Computer Communications - CSCI 551 Copyright © William C. ChengIf V(n) does not continue to increase as n goes to infinity,then we either need admission control or over-provisioningBest Effort never overloads (or does it?)11Utility Curve Shapes Computer Communications - CSCI 551 Copyright © William C. ChengBWUElasticBWUDelay-adaptiveBWUHard real-timeIf convex near origin, thenneed admission controlWorks for "normal users" because need to overprovision bya small amount12Over-provisioning Computer Communications - CSCI 551 Copyright © William C. ChengOver-provisioning for "leading edge" users is hard becausethese consume several orders of magnitude more than normalusersInternet will be provisioned to rarely block normal users, butwill block leading edge users frequentlyService should be explicitly requested by applicationsInternet should extend its service model13Summary Computer Communications - CSCI 551 Copyright © William C. ChengService model should incorporate admission controlData networkDigital convergency: Integrated ServicesTelephone networkAbstract formulation of maximizing efficacyCable network⇒ under one
View Full Document