57
MS I Networking Networking Kihong Park Kihong Park Assistant Professor Assistant Professor Department of Computer Sciences Department of Computer Sciences Purdue University Purdue University

MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

Embed Size (px)

Citation preview

Page 1: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

MSI

NetworkingNetworking

Kihong ParkKihong ParkAssistant ProfessorAssistant Professor

Department of Computer SciencesDepartment of Computer SciencesPurdue UniversityPurdue University

Page 2: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

22

MSI

TeamTeam

K. Park (CS)K. Park (CS) D. Yau (CS)D. Yau (CS) W. Szpankowski (CS)W. Szpankowski (CS) N. Shroff (ECE)N. Shroff (ECE) S. Fahmy (CS)S. Fahmy (CS)

Page 3: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

33

MSI

Related FacultyRelated Faculty

B. Bhargava (CS)B. Bhargava (CS) D. Comer (CS)D. Comer (CS) E. Houstis (CS)E. Houstis (CS) D. Marinescu (CS)D. Marinescu (CS) V. Rego (CS)V. Rego (CS)

Page 4: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

44

MSI

Funding (1)Funding (1)

Network SchedulingNetwork Scheduling• NSF ANI-9875789 (Park, CAREER)NSF ANI-9875789 (Park, CAREER)• NSF ANI-9714707 (Park)NSF ANI-9714707 (Park)

End System SchedulingEnd System Scheduling• NSF CCR-9875742 (Yau, CAREER)NSF CCR-9875742 (Yau, CAREER)• NSF ESS-9806741 (Grama, Park, Yau)NSF ESS-9806741 (Grama, Park, Yau)

Page 5: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

55

MSI

Funding (2)Funding (2)

RelatedRelated• NSF ANI-9729721 (vBNS, Purdue-wide)NSF ANI-9729721 (vBNS, Purdue-wide)• NSF NCR-9415491 (Szpankowski)NSF NCR-9415491 (Szpankowski)• NSF CCR-9804760 (Szpankowski)NSF CCR-9804760 (Szpankowski)• NSF NCR-9624525 (Shroff, CAREER)NSF NCR-9624525 (Shroff, CAREER)• NSF ANI-9805441 (Shroff, Schwartz)NSF ANI-9805441 (Shroff, Schwartz)• Sprint (Park, Spafford)Sprint (Park, Spafford)

Page 6: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

66

MSI

LabsLabs

Network Systems Lab (Park)Network Systems Lab (Park) Systems Software & Arch. Lab (Yau)Systems Software & Arch. Lab (Yau) NEWS (Shroff)NEWS (Shroff) Multimedia Compression (Szpankowski)Multimedia Compression (Szpankowski) BOND, COAST, Internetworking Lab, BOND, COAST, Internetworking Lab,

Multimedia DB, RAIDMultimedia DB, RAID

Page 7: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

77

MSI

ComponentsComponents

End-to-End QoSEnd-to-End QoS Network SchedulingNetwork Scheduling• Packet schedulingPacket scheduling• End-to-end controlEnd-to-end control

End System SchedulingEnd System Scheduling• Real-time OSReal-time OS• Distributed schedulingDistributed scheduling

Page 8: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

88

MSI

Research Goal (1)Research Goal (1)

Network EnvironmentNetwork Environment

Page 9: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

99

MSI

Research Goal (2)Research Goal (2)

Diverse QoS requirementsDiverse QoS requirements Shared network resourcesShared network resources

facilitate QoS-sensitive sharing of network facilitate QoS-sensitive sharing of network resourcesresources

Page 10: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1010

MSI

Research IssuesResearch Issues

Diverse user QoS requirementsDiverse user QoS requirements Bursty trafficBursty traffic ScalabilityScalability EfficiencyEfficiency Resource contention resolutionResource contention resolution SelfishnessSelfishness

Page 11: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1111

MSI

State-of-the-Art (1)State-of-the-Art (1)

Resource Reservation & Admission Resource Reservation & Admission ControlControl• On-demand leasing of bandwidthOn-demand leasing of bandwidth• QoS guaranteesQoS guarantees• GPS routersGPS routers• IETF Int-Serv (e.g., RSVP)IETF Int-Serv (e.g., RSVP)

Page 12: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1212

MSI

State-of-the-Art (2)State-of-the-Art (2)

DrawbacksDrawbacks• WastefulWasteful– self-similar bursty trafficself-similar bursty traffic– overprovisioningoverprovisioning

• Scales poorly with system sizeScales poorly with system size– per-flow controlper-flow control– admission control & policingadmission control & policing

Page 13: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1313

MSI

Page 14: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1414

MSI

State-of-the-Art (3)State-of-the-Art (3)

Overkill for many applicationsOverkill for many applications• Soft QoSSoft QoS• Graded service levelsGraded service levels

Resource contention resolutionResource contention resolution• ``Who should get what’’``Who should get what’’• Noncooperative environmentNoncooperative environment

Page 15: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1515

MSI

State-of-the-Art (4)State-of-the-Art (4)

IETF Diff-ServIETF Diff-Serv• Assured service (Clark)Assured service (Clark)

2-bit marking bit RIO gateways2-bit marking bit RIO gateways

• Premium service (Jacobson)Premium service (Jacobson) Leaky-bucket with priority gatewaysLeaky-bucket with priority gateways

aggregate flow control with weak aggregate flow control with weak protectionprotection

Page 16: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1616

MSI

State-of-the-Art (5)State-of-the-Art (5)

DrawbacksDrawbacks• ``Differentiated’’ service``Differentiated’’ service

not QoS-sensitivenot QoS-sensitive

• Admission control & policing/shapingAdmission control & policing/shaping impediment to scalabilityimpediment to scalability

• Weak complexity trade-offWeak complexity trade-off• Lack of resource contention resolutionLack of resource contention resolution

Page 17: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1717

MSI

SBS Approach (1)SBS Approach (1)

Dual ArchitectureDual Architecture• Guaranteed service (GS)Guaranteed service (GS)– Reservation-based QoSReservation-based QoS– Small fraction of most stringent appsSmall fraction of most stringent apps

• Stratified best-effort service (SBS)Stratified best-effort service (SBS)– Graded serviceGraded service– Large spectrum of QoS-sensitive appsLarge spectrum of QoS-sensitive apps

Page 18: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1818

MSI

SBS Approach (2)SBS Approach (2)

Architectural FeaturesArchitectural Features• Compatible with reservation schemeCompatible with reservation scheme– Common substrate of GPS routersCommon substrate of GPS routers– Amortize cost of overprovisioning for GSAmortize cost of overprovisioning for GS

• Aggregate flow controlAggregate flow control– Per-flow QoS controlPer-flow QoS control– Zero per-flow state at routersZero per-flow state at routers

• Distributed protocolDistributed protocol

Page 19: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

1919

MSI

SBS Approach (3)SBS Approach (3)

SBS RouterSBS Router

Page 20: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2020

MSI

SBS Approach (4)SBS Approach (4)

SBS Control PathSBS Control Path

Page 21: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2121

MSI

SBS Approach (5)SBS Approach (5)

PropertiesProperties• Graded servicesGraded services– Beyond differentiated servicesBeyond differentiated services– QoS-sensitivityQoS-sensitivity

• ScalabilityScalability– No resource reservationNo resource reservation– No admission controlNo admission control– Aggregate flow controlAggregate flow control

• Simple user/network interfaceSimple user/network interface

Page 22: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2222

MSI

SBS Approach (6)SBS Approach (6)

• Resource contention resolutionResource contention resolution– Selfish user behaviorSelfish user behavior– Stability, optimality, efficiencyStability, optimality, efficiency– PricingPricing

• WAN implementationWAN implementation– User-centric User-centric networking-centric mechanism networking-centric mechanism– Simple user/simple network realizationSimple user/simple network realization– Distributed QoS ControlDistributed QoS Control

Page 23: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2323

MSI

Distributed QoS Control (1)Distributed QoS Control (1)

Component ServicesComponent Services• Guaranteed service (GS)Guaranteed service (GS)• Stratified best-effort service (SBS)Stratified best-effort service (SBS)

SBS ProtocolSBS Protocol• Packet scheduling at routersPacket scheduling at routers• End-to-end controlEnd-to-end control

Page 24: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2424

MSI

Distributed QoS Control (2)Distributed QoS Control (2)

WAN QoS Control: GSWAN QoS Control: GS

Page 25: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2525

MSI

Distributed QoS Control (3)Distributed QoS Control (3)

WAN QoS Control: SBS + GSWAN QoS Control: SBS + GS

Page 26: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2626

MSI

Distributed QoS Control (4)Distributed QoS Control (4)

Self-Optimization ProcedureSelf-Optimization Procedure Select service class s.t.Select service class s.t.1 . QoS requirement is satisfied. QoS requirement is satisfied2 . Least cost, i.e., minimum resource service . Least cost, i.e., minimum resource service

classclass

• User optimalUser optimal• Selfishness emulationSelfishness emulation

Page 27: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2727

MSI

Distributed QoS Control (5)Distributed QoS Control (5)

GPS Switch & Multi-Class QoS GPS Switch & Multi-Class QoS

Page 28: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2828

MSI

Distributed QoS Control (6)Distributed QoS Control (6)

Service Class-QoS Association TableService Class-QoS Association Table

CCa a << CCb b PPa a >> PPbb

Page 29: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

2929

MSI

System Operation (1)System Operation (1)

Service Access PointService Access Point• Direct user accessDirect user access• Access by DB & Security modulesAccess by DB & Security modules

Export APIExport API (GS/SBS, QoS Requirement Vector)(GS/SBS, QoS Requirement Vector)

Page 30: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3030

MSI

System Operation (2)System Operation (2)

Interface with DB & Security ModulesInterface with DB & Security Modules

Page 31: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3131

MSI

QoS Parameters & Translation (1)QoS Parameters & Translation (1)

Network QoSNetwork QoS• Traffic characterizationTraffic characterization– Mean/peak/minimum data ratesMean/peak/minimum data rates– Reliable/unreliable transportReliable/unreliable transport

• QoS requirementsQoS requirements– Packet loss ratePacket loss rate– DelayDelay– JitterJitter

Page 32: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3232

MSI

QoS Parameters & Translation (2) QoS Parameters & Translation (2)

Application QoSApplication QoS QoP & other user requirementsQoP & other user requirements– Video frame rate (f/s)Video frame rate (f/s)– Hit rateHit rate– Hit rate jitterHit rate jitter– ReliabilityReliability– Security servicesSecurity services

Page 33: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3333

MSI

QoS Parameters & Translation (3)QoS Parameters & Translation (3)

Hit RateHit Rate

Probability of timely arrivalProbability of timely arrival

Page 34: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3434

MSI

QoS Parameters & Translation (4)QoS Parameters & Translation (4)

QoS TranslationQoS Translation

Page 35: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3535

MSI

QoS Parameters & Translation (5)QoS Parameters & Translation (5)

Example (hit rate)Example (hit rate)

if Measured Hit Rate < Target Hit Rate thenif Measured Hit Rate < Target Hit Rate then

increase allocated resourcesincrease allocated resources

elseelse

decrease resourcesdecrease resources

QoS translation & control loopQoS translation & control loop

Page 36: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3636

MSI

System Performance (1)System Performance (1)

vBNS BenchmarkingvBNS Benchmarking

Page 37: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3737

MSI

System Performance (2)System Performance (2)

Traffic ConfigurationTraffic Configuration

Page 38: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3838

MSI

System Performance (3)System Performance (3)

SBS PerformanceSBS Performance

Page 39: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

3939

MSI

System Performance (4)System Performance (4)

Performance ComparisonPerformance Comparison

Page 40: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4040

MSI

End-to-End QoS Control (1)End-to-End QoS Control (1)

SBS Exports Graded ServicesSBS Exports Graded Services

QoS amplification using end-to-end controlQoS amplification using end-to-end control

Page 41: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4141

MSI

End-to-End QoS Control (2)End-to-End QoS Control (2)

Real-Time TrafficReal-Time Traffic Adaptive forward error-correctionAdaptive forward error-correction

• Inject redundancy adaptively (network state)Inject redundancy adaptively (network state)• Maintain invariant target QoSMaintain invariant target QoS• Optimal control problemOptimal control problem• AFEC protocolAFEC protocol

Page 42: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4242

MSI

End-to-End QoS Control (3)End-to-End QoS Control (3)

• QoS-Redundancy CurveQoS-Redundancy Curve

Page 43: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4343

MSI

End-to-End QoS Control (4)End-to-End QoS Control (4)

• Real-Time MPEG Video TransportReal-Time MPEG Video Transport

Page 44: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4444

MSI

End-to-End QoS Control (5)End-to-End QoS Control (5)

• Benchmark ExperimentsBenchmark Experiments

Page 45: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4545

MSI

End-to-End QoS Control (6)End-to-End QoS Control (6)

• Benchmark DataBenchmark Data

Page 46: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4646

MSI

End-to-End QoS Control (7)End-to-End QoS Control (7)

• Performance MeasurementsPerformance Measurements

Page 47: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4747

MSI

End-to-End QoS Control (8)End-to-End QoS Control (8)

• Redundancy TraceRedundancy Trace

Page 48: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4848

MSI

End-to-End QoS Control (9)End-to-End QoS Control (9)

Self-Similar Traffic ControlSelf-Similar Traffic Control Exploit long-range dependenceExploit long-range dependence

• Multiple time scale congestion control (MTSC)Multiple time scale congestion control (MTSC)• Throughput (SAC, TCP-MT)Throughput (SAC, TCP-MT)• Real-time traffic (AFEC-MT)Real-time traffic (AFEC-MT)

Page 49: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

4949

MSI

End-to-End QoS Control (10)End-to-End QoS Control (10)

• Multiple time scale congestion controlMultiple time scale congestion control

Page 50: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5050

MSI

Proposed Research (1)Proposed Research (1)

Implementation of GS/SBS ArchitectureImplementation of GS/SBS Architecture• Experimental network Experimental network

``Purdue Infobahn’’``Purdue Infobahn’’ Testbed & substrate for MSITestbed & substrate for MSI

• User/network APIUser/network API• 2-Tier network2-Tier network– Backbone: Cisco 72xx routersBackbone: Cisco 72xx routers– LANs: ATM & Gigabit EthernetLANs: ATM & Gigabit Ethernet

Page 51: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5151

MSI

Proposed Research (2)Proposed Research (2)

• Purdue InfobahnPurdue Infobahn

Page 52: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5252

MSI

Proposed Research (3)Proposed Research (3)

Performance Comparison: Diff-ServPerformance Comparison: Diff-Serv• Assured serviceAssured service• Premium servicePremium service• SBS OptimizationSBS Optimization– End-to-end QoS control End-to-end QoS control switch architecting switch architecting– Cooperative controlCooperative control– Refinements (e.g., SBS Light)Refinements (e.g., SBS Light)

• Deployability experimentsDeployability experiments

Page 53: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5353

MSI

Proposed Research (4)Proposed Research (4)

Measurement Collection/DisseminationMeasurement Collection/Dissemination• Traffic characterizationTraffic characterization• User behavior characterizationUser behavior characterization– INDEX project (Berkeley, NSF ANIR)INDEX project (Berkeley, NSF ANIR)– Limitations:Limitations:• Outdated applicationsOutdated applications• ISDN line ratesISDN line rates• Infinite supply/overprovisioningInfinite supply/overprovisioning

Page 54: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5454

MSI

Proposed Research (5)Proposed Research (5)

• User behavior characterization (cont.) User behavior characterization (cont.) – Next generation resource sharing behaviorNext generation resource sharing behavior• high-bandwidth, multimedia applicationshigh-bandwidth, multimedia applications

extend to direct user poolextend to direct user pool• diverse QoS requirementsdiverse QoS requirements• token/pricing-based resource contention resolutiontoken/pricing-based resource contention resolution• stability, robustness, fairnessstability, robustness, fairness

• vBNS/Abilene peering trafficvBNS/Abilene peering traffic

Page 55: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5555

MSI

Proposed Research (6)Proposed Research (6)

End-to-End QoS Control ExperimentsEnd-to-End QoS Control Experiments• QoS amplification trade-offQoS amplification trade-off

``Quality vs. quantity’’``Quality vs. quantity’’ QoS tunnelingQoS tunneling Protocol stack interactionProtocol stack interaction

• Self-similar traffic controlSelf-similar traffic control Long-range dependence & delay-bandwidthLong-range dependence & delay-bandwidth

• Variable compression & traffic controlVariable compression & traffic control

Page 56: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5656

MSI

Proposed Research (7)Proposed Research (7)

Integrated DB & Security ExperimentsIntegrated DB & Security Experiments• Extended QoS specification & APIExtended QoS specification & API• QoS translationQoS translation• Integrated schedulingIntegrated scheduling– Component analysisComponent analysis– Hybrid schedulingHybrid scheduling

• Application benchmarkingApplication benchmarking

Page 57: MS I Networking Kihong Park Assistant Professor Department of Computer Sciences Purdue University

5757

MSI

Proposed Research (8)Proposed Research (8)

QoS, Security, and Fault-ToleranceQoS, Security, and Fault-Tolerance• Security/QoS trade-off: overheadSecurity/QoS trade-off: overhead

``Reactive security’’ (vs. proactive)``Reactive security’’ (vs. proactive)

• Fault-tolerance and QoSFault-tolerance and QoS– Impact of fault-tolerance on QoSImpact of fault-tolerance on QoS– Hierarchical fault-tolerance & self-healingHierarchical fault-tolerance & self-healing– Network managementNetwork management– AvailabilityAvailability