The dissertation of Suchitra Raman is approved

140 Pages·2000·1.26 MB·English

AFrameworkforInteractive MulticastDataTransport inthe Internet
by
SuchitraRaman
B.Tech. (IndianInstituteofTechnology,Madras,India)1996
M.S.(UniversityofCalifornia,Berkeley) 1998
Adissertationsubmittedinpartial satisfactionofthe
requirementsforthedegreeof
DoctorofPhilosophy
in
ComputerScience
inthe
GRADUATEDIVISION
ofthe
UNIVERSITY ofCALIFORNIA at BERKELEY
Committeeincharge:
ProfessorStevenR.McCanne,Chair
ProfessorRandyH.Katz
ProfessorKennethGoldberg
2000
ThedissertationofSuchitraRamanisapproved:
Chair Date
Date
Date
UniversityofCaliforniaat Berkeley
2000
AFrameworkforInteractive MulticastDataTransport inthe Internet
Copyright2000
by
SuchitraRaman
1
Abstract
AFrameworkforInteractiveMulticastDataTransportintheInternet
by
SuchitraRaman
DoctorofPhilosophyinComputerScience
UniversityofCaliforniaat Berkeley
ProfessorStevenR.McCanne,Chair
Theremarkable growth of the Internet as the a data transmission medium has in part been enabled
by the simplicity and scalability of the core Internet Protocol (IP), which is used for addressing
and routing unicast data packets through the network. The IP service model does not provide any
packet delivery guarantees, but rather provides a “best-effort” contract, and leaves it to higher lay-
ers toprovide enhanced services using this basic service. Today, the defacto transport protocol on
the Internet is the Transmission Control Protocol (TCP) [109, 128]. TCP was designed primarily
for applications such as telnet, a remote terminal application, and ftp, a file transfer application,
which require data to be delivered reliably and in an ordered manner. While the TCP abstraction
and protocol are indeed invaluable for transporting simple data types in telnet and ftp, as well as
otherapplicationswithstraightforward reliabilityrequirements, twonotablechangesmaketheTCP
approach inappropriate both with respect to its restrictive delivery semantics as well as its internal
algorithms for loss recovery. First, TCP is not compatible with extensions to the basic IP service
modelforprovidingnetwork-layermulticast. Second,theemergenceofrichmediatypesandappli-
cations creates a need for transport protocol requirements that are not satisfied by TCP’srestricted
semantics. The newer applications are interactive and handle these special media types in special
ways. Forexample,animageviewerthatdoesnotrelyonTCPmayhandleJPEGimagedatadeliv-
eredoutoforderandreconstructmissingportionsusinginterpolationtechniques,therebyenhancing
theinteractivity totheend user. Suchapplications require sophisticated delivery semantics andare
not best served by an overly restrictive protocol such as TCP. What is required here is a transport
protocol whosesemanticscanbetailored bytheapplication forefficientnetworktransmission.
Ourapproach tosolving theseissues isasoftstate-based interactive multicast datatrans-
portprotocolframework. Wepresentamodelfor“softstate”asanend-to-endconstructthatenables
loosestatesynchronization betweensenderandreceivers. Wetreatprotocol control stateattheend
points as “soft” by not requiring that it be perfectly consistent at all instants. This allows us to
avoid tight sender-receiver synchronization, as in TCP-like instantaneous receiver acknowledge-
ments. Our soft state-based transport protocol provides a relaxed reliability, instead of TCP-like
deterministic reliability.
Second, to accommodate heterogeneity among receivers and network paths, we allow
receivers to tailor the semantics of reliability. Hence, a receiver incapable of or uninterested in
processing portions ofthedatastreammayrefrainfrom receiving itreliably. Wedonotrelyonthe
transport-level sequence space, but rather, use application-specific namespaces to express receiver
2
preferences while requesting retransmissions. This application-level namespace is exposed to the
transport protocol and isused by the receiver toselectively retrieve specific data items. Theuse of
suchacommon“vocabulary” todescribedataputstheapplication incontroloflossrecovery.
Finally,sincemanynewdatatypesincludingcertainimageformatscanbeprocessed and
renderedoutoforderatthereceiver,wedonotenforceaTCP-likedeliveryorderonthedatastream.
Instead, we provide out-of-order delivery to the receiving application and demonstrate its benefits
forimagedelivery. Thisspecifictechnique isalsoapplicable tounicasttransmission andwedesign
andimplementaninteractive imagetransmission protocol foruseintheWorldWideWeb.
These techniques form the bases of the new transport protocol framework for interactive
multicast data transport. Our transport protocol is layered on top of UDP [108] in the protocol
stack,andwehaveimplementeditasauser-levellibrarycalledlibsstp,alibraryforsoftstate-based
transport protocol. We also present probabilistic analyses of the performance of our protocol in
termsoftheperformanceofthebasicalgorithmsforlossrecovery,using“slottinganddamping,”as
wellasthetradeoffsinvolving consistency andbandwidth consumption.
ProfessorStevenR.McCanne
Dissertation CommitteeChair
iii
Tomyparents,LakshmiandHariharRaman
iv
Contents
ListofFigures vii
ListofTables ix
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 TheProblem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 OurSolution: Interactive MulticastTransportUsingSoftState . . . . . . . . . . . 6
1.3.1 SoftStateasaDataTransportPrimitive . . . . . . . . . . . . . . . . . . . 7
1.3.2 Receiver-driven Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.3 Out-of-Order Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Contributions ofDissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 OverviewofDissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2 BackgroundandRelatedWork 13
2.1 OverviewofIPMulticast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 FloodandPruneProtocols . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2 ExplicitTreeProtocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Multicast Deployment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Announce/Listen-based TransportProtocols . . . . . . . . . . . . . . . . . . . . . 17
2.4 ReliableMulticastTransportProtocols . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4.1 ScalableReliableMulticast(SRM) . . . . . . . . . . . . . . . . . . . . . 18
2.4.2 Tree-based ReliableMulticast Protocols . . . . . . . . . . . . . . . . . . . 20
2.4.3 Router-assisted Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.4 ReliableMulticastFramework(RMF) . . . . . . . . . . . . . . . . . . . . 21
2.4.5 ForwardErrorCorrection-based Schemes . . . . . . . . . . . . . . . . . . 21
2.5 DeliverySemanticsinTransportProtocols . . . . . . . . . . . . . . . . . . . . . . 21
2.6 SummaryofRelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 SoftState-basedTransport 24
3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 TheDataModel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.1 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 “Open-Loop” Announce/Listen Protocol . . . . . . . . . . . . . . . . . . . . . . . 28
v
3.4 MultipleTransmission Queues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5 ImpactofReceiverFeedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.6 ASoftStateTransportFramework . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6.1 Application-controlled BandwidthAllocation . . . . . . . . . . . . . . . . 37
3.6.2 Hierarchical DataModel . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.7 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4 ScalableDataNaming 45
4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2 SNAP:Hierarchical DataNaming . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 SourceIdentifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.4 Discovering theSNAPNamespace . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4.1 Data-driven LossDetection . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.4.2 Announcement-driven LossDetection . . . . . . . . . . . . . . . . . . . . 53
4.5 SNAP:PerformanceEvaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.6 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.7 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5 AsymptoticScalingofRandomizedTimers 62
5.1 OverviewofRandomized Timers . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2 PreviousWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3 SimulationMethodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.4 ScalingintheConeTopology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.5 ScalingintheLinearChain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.5.1 ExactRTTEstimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.5.2 WithoutRTTEstimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.6 ScalingintheBinaryTree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.7 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6 ITP:AnImageTransportProtocol 82
6.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.2 DesignConsiderations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.3 ITPDesign . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.3.1 Out-of-order Delivery . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.3.2 Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3.3 UsingtheCongestion Manager . . . . . . . . . . . . . . . . . . . . . . . . 92
6.3.4 DesignSummary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.4 JPEGTransportusingITP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.4.1 Framing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.4.2 Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.4.3 ErrorConcealment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.4.4 OtherFormats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.5.1 PeakSignal-to-Noise Ratio(PSNR) . . . . . . . . . . . . . . . . . . . . . 96
vi
6.5.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7 libsstp: AUser-levelTransportProtocolforInteractive MulticastApplications 102
7.1 LibsstpSoftwareArchitecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.1.1 SessionObject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.1.2 SourceObject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.1.3 DataPath . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.1.4 Application Callbacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.1.5 EventHandling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.2 LibsstpApplications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8 ConclusionsandFutureWork 112
8.1 FutureDirections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.1.1 SoftStateModelforRSVP . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.1.2 CompactNamespaceRepresentations . . . . . . . . . . . . . . . . . . . . 113
8.1.3 ContentPeering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.1.4 NapsterOverlayNetworks . . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.1.5 Hierarchical SessionDirectory . . . . . . . . . . . . . . . . . . . . . . . . 114
8.1.6 Multicast-based SoftwareUpdates . . . . . . . . . . . . . . . . . . . . . . 115
8.2 Availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Bibliography 117
vii
List of Figures
1.1 Thefeedbackimplosion problem inmulticastsessions. . . . . . . . . . . . . . . . 4
1.2 Graphshowingthattightsynchronization doesnotscalegracefully withgroupsize. 5
1.3 Protocolstackshowinglibsstp. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 TCPadversely affectsimagedownloadperformance. . . . . . . . . . . . . . . . . 10
1.5 Negativeimpactofin-orderdeliveryonimagedownloads. . . . . . . . . . . . . . 10
2.1 Multicast routingandgroupmembership. . . . . . . . . . . . . . . . . . . . . . . 14
2.2 NACKsuppression inScalableReliableMulticast. . . . . . . . . . . . . . . . . . 19
3.1 Thesoftstatedatamodel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Queueing modelforannounce/listen-based transport protocol. . . . . . . . . . . . 28
3.3 Impactofpacketlossandannouncement deathrateontheconsistency metric. . . . 30
3.4 Bandwidthoverheadinthesoftstaterefreshprotocol. . . . . . . . . . . . . . . . . 31
3.5 Performance oftwo-levelscheduling onsoftstateprotocol performance. . . . . . . 32
3.6 Theeffectofbandwidth allocation onlatency. . . . . . . . . . . . . . . . . . . . . 33
3.7 Statediagram showingtheevolution ofadataitematthesender. . . . . . . . . . . 34
3.8 Performance oftheopenloopprotocol. . . . . . . . . . . . . . . . . . . . . . . . 35
3.9 Effectofreceiverfeedback onsoftstateprotocol performance. . . . . . . . . . . . 36
3.10 Tradeoffsinbandwidthallocation atthesender. . . . . . . . . . . . . . . . . . . . 37
3.11 Optimalbandwidth allocation basedonprotocol consistency. . . . . . . . . . . . . 38
3.12 Profile-driven scheduler forthesoftstateprotocol. . . . . . . . . . . . . . . . . . . 39
3.13 Theeffectofannouncement frequency andNACKbandwidthonSSTPperformance. 40
3.14 Theeffectofinputdatarateontheaverageconsistency metricinthemulticastcase. 41
3.15 Theeffectofobjectgranularity onsoftstateprotocol performance. . . . . . . . . . 42
4.1 AnexampleofaSNAPnamespace. . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Therandomizedinitialnodedescriptor (IND)inSNAP. . . . . . . . . . . . . . . . 50
4.3 Namemapbindings intheSNAPnamespacehierarchy. . . . . . . . . . . . . . . . 52
4.4 Recovering fromtaillossesusingannouncements. . . . . . . . . . . . . . . . . . . 56
4.5 Retrieving missingnodesusingSNAPnamespaceannouncements. . . . . . . . . . 57
4.6 Tuningreliability usingSNAP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.7 Convergence behavior ofSNAPwithdecreasing frequency ofupdates. . . . . . . . 59
4.8 Effectiveness ofSRM-stylesuppression inSNAP. . . . . . . . . . . . . . . . . . . 60
4.9 Softwarearchitecture oftheSNAPimplementation. . . . . . . . . . . . . . . . . . 61