EECS 454EECS 454: Advanced Communication NetworksSpring Quarter 2010Meeting time: 3:30-4:50 MWInstructor: Randall BerryOffice: Tech, Rm. M318Office Hours: by appointmentEECS 454Course OverviewPrimary goal is to develop analytical tools and conceptualmodels that are useful in networking research (and in otherfields).Focus of the course is on issues identified with data link,network and transport layer - physical layer and applicationlayer issues will not be addressed in detail.This graduate level course - you should question what youare learning and think about how what you are learningmay be applied in other situations. What are thelimitations, key assumptions, etc.EECS 454Course InformationPrerequisites:Good understanding of basic probability. (If you are notcomfortable with probability, it may be helpful to take EECS422 before taking this course.)Familiarity with data networks (e.g. EECS 333 or 340) ishelpful (this provides motivation/context for the materialstudied here).Text: Data Networks, 2nd Ed. by D. Betsekas and R. Gallager.For the first half of the course we will follow parts the textclosely; for the second half, the text will be supplementedwith other notes and journal papers. Other references arelisted on the course information sheet.EECS 454GradingProblem Sets - can work in groups - write-up your ownsolutions.1st mid-term - in-class, about 1/2 way through term.2nd mid-term - take home exam - last week of classesFinal Project - presentation and write-up during last weekof classes or final exam week.EECS 454EECS 454 vs. EECS 333Fewer topics than in 333, but cover each topic in moredepth.Less emphasis than 333 on describing actual protocolsand implementation issues, more emphasis on analyticaltechniques and performance issues.EECS 454Research in NetworkingMultidisciplinary/multifaceted.The material in this course focuses on the analytical side.But this far from the whole story, also a lot of work onmeasurements, simulation, implementation, protocoldesign, etc.EECS 454Analytical ModelingMost of this course consists of describing different analyticaltechniques that can be used for understanding various aspectsof networking.Uses:Performance analysisParameter setting, network provisioning, comparison ofdifferent approaches.Improve understanding/intuitionqualitative behavior, performance trade-offs, systembottlenecks.Establish fundamental limitationsWhat is best performance possible?EECS 454Analytical ModelingImportant point to remember: All models are wrong.So what makes a good model?EECS 454Analytical ModelingTwo uses:1Gain deeper insight into performanceOften simplified “toy” models - need to abstract away muchdetail to get something you can really understand.Need a good understanding of what is being modeled andof the modeling techniques used.Often developing such models is an iterative process.2Performance analysisMore detailed - want to acurately capture actual systemperformance- often too complicated to analyze by hand.EECS 454Communication NetworksPhysically communication networks consist of“communication links” connecting together different“nodes” for the purpose of exchanging information (bits).Here focus mainly on “wire-line” network models(point-to-point links).Each link can be thought of as a (possibly lossy) “bit-pipe”with a certain transmission rate.Nodes - sources, sinks, routers/switches.Mathematically - represent topology as a graph.EECS 454Networking motivationsNetworks are used to provide different information servicesto end-users (e.g. e-mail, ftp, voice telephony, streamingvideo,..).Key reason for using networks is resource sharing(economies of scale/economies of scope).Different applications require different qualities of service(QoS).A key issue is how to share resources and satisfy QoS.EECS 454Packets, Sessions, etc.“Users” initiate “sessions” and exchange “messages”.connection-oriented/connectionlessMessages may be broken into smaller “packets” to be sentover network.Often special names used in particular contexts: frames,cells, ....EECS 454MultiplexingMultiplexing refers to techniques for sharing a link amoungdifferent sessions.Two main techniquescircuit-based: e.g. TDM/FDM.Performance: blocking probability.Packet-based: statistical multiplexing/scheduling.performance: delay/packet dropping.Various hybrids also possible.EECS 454Switching/routingA related issue to mulitplexing is how packets are switched androuted within the network.Routing: how to determine the next link for a packet.Switching: how nodes physically send packets from onelink to the next.EECS 454Network ControlAnother key issue is how to “control” the traffic in thenetwork.A variety of possible techniques:Flow control, congestion control, admission control..EECS 454Network Layers and ProtocolsConceptually, we think of networks as a sequence ofvertical layers; each layer providing some type of service tolower layers.Higher layers provide a higher level abstraction of thenetwork.Each layer provides its service by implementing one ormorecommunication protocols.A protocol specifies how different processes in a networkinteract; this includes specifying the format of messagesthat are exchanged and the algorithm used to generatethese messages.A fundamental characteristic of most of the algorithms usedin networks is that they are distributed.EECS 454Perfomance AnalysisThe two most common performance metrics used in networkingareDelay - i.e., how long does it take to send information fromits source to its destination.Throughput - i.e. how much data per second can be sentacross the network.Much of the behaviour in networks is best modeled as random(e.g. user behaviour, failures); thus performance analysis istypically done in a probabalistic framework.EECS 454Modleing scalesTraffic in a network can be modeled at different scales.At the finest scale we have the arrivals and departures ofindividual packets in a session.Such models are refered to as ”packet-level models.”At a larger scale, one can look at ”flow level models.”Focus on the arrival and departure of flows as the randomquantity of interest.Often model the transmission of packets within a flow as a”fluid” process.Can think of as looking at system on a longer time-scale.In some cases, it is even reasonable to focus on a fixed setof active flows.This leads to deterministic models, e.g. diferential equationmodels and optimization-based models of TCP.EECS 454Course OutlineQuick