1. 1 Performance Study of Congestion Price Based Adaptive Service Performance Study of Congestion Price Based Adaptive Service Xin Wang, Henning Schulzrinne Xin Wang, Henning Schulzrinne (Columbia University) http://www.cs.columbia.edu/xinwang/public /projects/RNAP.html

2. 2Motivation Combine resource reservation with multimedia adaptive serviceCombine resource reservation with multimedia adaptive service Pricing network services based on level of service, usage, and congestion - a natural and fair incentive for applications to adapt their sending rates according to network conditions.Pricing network services based on level of service, usage, and congestion - a natural and fair incentive for applications to adapt their sending rates according to network conditions. Our work:Our work: –resource commitment short interval –trade-offs between blocking and raising prices in network

3. 3 Outline Resource negotiation & RNAPResource negotiation & RNAP Pricing strategyPricing strategy User adaptationUser adaptation Simulation modelSimulation model Results and discussionResults and discussion RNAP message aggregationRNAP message aggregation ConclusionConclusion

4. 4 Resource Negotiation & RNAP Assumption: network provides a choice of delivery services to userAssumption: network provides a choice of delivery services to user –e.g. diff-serv, int-serv, best-effort, with different levels of QoS –with a pricing structure (may be usage- sensitive) for each. RNAP (Resource Negotiation and Pricing): a protocol through which the user and network (or two network domains) negotiate network delivery services.RNAP (Resource Negotiation and Pricing): a protocol through which the user and network (or two network domains) negotiate network delivery services. –Network -> User: communicate availability of services; price quotations and accumulated charges –User -> Network: request/re-negotiate specific services for user flows. Underlying Mechanism: combine network pricing with traffic engineeringUnderlying Mechanism: combine network pricing with traffic engineering

5. 5 Resource Negotiation & RNAP (cont’d.) Who can use RNAP?Who can use RNAP? –Adaptive applications: adapt sending rate, choice of network services –Non-adaptive applications: take fixed price, or absorb price change

6. 6 Protocol Architectures: Centralized Architecture (RNAP-C) NRN S1S1 R1R1 HRN Access Domain - B Access Domain - A Transit Domain Internal Router Edge Router Host RNAP Messages NRN HRN Network Resource Negotiator Host Resource Negotiator Intra domain messages Data

7. 7 S1S1 R1R1 HRN Access Domain - B Access Domain - A Transit Domain Protocol Architectures: Distributed Architecture (RNAP-D) Internal Router Edge Router Host RNAP Messages NRN HRN Intra domain messages Data

8. 8 RNAP messages Query Quotation Reserve Commit Quotation Reserve Commit Close Release Query: User enquires about available services, prices Quotation: Network specifies services supported, prices Reserve: User requests service(s) for flow(s) (Flow Id- Service-Price triplets) Commit: Network admits the service request at a specific price or denies it (Flow Id- Service-Status-Price) Periodic re-negotiation Close: tears down negotiation session Release: releases the resources

9. 9 Pricing on Current Internet Access rate dependent charge (AC)Access rate dependent charge (AC) Volume dependent charge (V)Volume dependent charge (V) AC + V AC-VAC + V AC-V Usage based charging: time-based / volume- basedUsage based charging: time-based / volume- based

10. 10 Pricing Strategy Fixed pricingFixed pricing –Service class independent flat pricing –Service class sensitive priority pricing –Time dependent time of day pricing –Time-dependent service class sensitive priority pricing

11. 11 Pricing Strategy (cont’d.) Congestion-based PricingCongestion-based Pricing –Usage charge: p u = f (service, demand, destination, time_of_day,...) c u (n) = p u x V (n) –Holding charge: P h i =  i x (p u i - p u i-1 ) c h (n) = p h x R(n) x  –Congestion charge: p c (n) = min [{p c (n-1) +  (D, S) x (D- S)/S,0 } +, p max ] c c (n) = p c (n) x V(n)

12. 12 User Adaptation Based on perceived valueBased on perceived value Maximize total utility over the total costMaximize total utility over the total cost Constraint: budget, min QoS & max QoSConstraint: budget, min QoS & max QoS

13. 13 User Adaptation (cont’d.) An example utility functionAn example utility function –U (x) = U 0 +  log (x / x m ) Optimal user demandOptimal user demand – Without budget constraint: x j =  j / p j –With budget constraint: x j = (b x  j / Σ l  l ) / p j Affordable resource is distributed proportionally among applications of the system, based on the user’s preference and budget for each application.Affordable resource is distributed proportionally among applications of the system, based on the user’s preference and budget for each application.

14. 14 Simulation Model Topology 1 Topology 2

15. 15 Simulation Model (cont’d.) Parameter Set-upParameter Set-up –topology1: 48 users –topology 2: 360 users –user requests: 60 kb/s -- 160 kb/s –targeted reservation rate: 90% –price adjustment factor: σ = 0.06 –price update threshold: θ = 0.05 –negotiation period: 30 seconds –usage price: p u = 0.23 cents/kb/min –average session length 10 minutes, exponential distributed. Performance measuresPerformance measures –Bottleneck bandwidth utilization –User request blocking probability –Average and total user benefit –Network revenue –System price –User charge

16. 16 Design of the Experiments Performance comparison of congestion-based pricing system (CPA) with a fixed-price based system (FP)Performance comparison of congestion-based pricing system (CPA) with a fixed-price based system (FP) Effect of system control parameters:Effect of system control parameters: – target reservation rate –price adjustment step –price adjustment threshold Effect of user demand elasticityEffect of user demand elasticity Effect of session multiplexingEffect of session multiplexing Effect when part of users adaptEffect when part of users adapt Session adaptation and adaptive reservationSession adaptation and adaptive reservation

17. 17 Bottleneck Utilization Request blocking probability

18. 18 Total network revenue ($/min) Total user benefit ($/min)

19. 19 Price ($/kb/min) User bandwidth (kb/s)

20. 20 Average user charge ($/min)

21. 21 Bottleneck utilization Request blocking probability Effect of target utilization level

22. 22 Effect of Price Adjustment Step Request blocking probability Bottleneck utilization

23. 23 Effect of Price Adjustment Threshold Bottleneck utilization Request blocking probability

24. 24 Effect of User Demand Elasticity Average user bandwidth Average user charge

25. 25 Effect of Multiplexing Between User Sessions Request blocking probability Total user benefit

26. 26 Adaptation by Part of User Population Bottleneck utilization Request blocking probability

27. 27 One-time Versus Ongoing Adaptation Bottleneck utilization Request blocking probability

28. 28 RNAP Message Aggregation RNAP-D RNAP-C

29. 29 RNAP Message Aggregation (cont’d) Aggregate for senders sharing the same destination networkAggregate for senders sharing the same destination network –merged by source domains –split for HRNs at destination net: border router (RNAP-D) NRN (RNAP-C) Two messages:Two messages: –aggregated-resource message reserves and collects price in the middle of network –original messages sent directly to destination without visiting agents in between

30. 30 Conclusions CPA gain over FPCPA gain over FP –Network availability, revenue, user perceived benefit –CPA congestion control is stable and effective Target reservation rate (utilization):Target reservation rate (utilization): –too high or too low utilization: User benefit –Too low target rate: demand fluctuation is high – Too high target rate: high blocking rate

31. 31 Conclusions Effect of price scaling factor Effect of price scaling factor  – , blocking rate –  too large under-utilization, large dynamics Effect of price adjustment threshold Effect of price adjustment threshold  –Too high, no meaningful adaptation

32. 32 Conclusions Demand elasticityDemand elasticity –Bandwidth sharing is proportional to user’s willingness to pay User adaptation by some users still results in overall performance improvementUser adaptation by some users still results in overall performance improvement