PRICING FOR MOBILE DATA Sangtae Ha Princeton University Joint work - PowerPoint PPT Presentation

TUBE: TIME DEPENDENT PRICING FOR MOBILE DATA Sangtae Ha Princeton University Joint work with: Soumya Sen, Carlee Joe-Wong, Youngbin Im, Mung Chiang

2 (Mobile) Data Explosion Mobile data growing at 78% annually 2.4 billion mobile users worldwide, 260 million US mobile data users by 2017

3 Driving Forces Mobile Cloud Data-hungry High-res A Perfect Video Sync Apps Devices Storm

4 Industry Moves: US ISPs Wireless Elimination of Unlimited Data Plans Introduction of Usage-based penalties Verizon eliminates new No unlimited plans unlimited smartphone plans offered for iPad LTE (July 2011) 8 (March 2012) 10 AT&T throttles Verizon to eliminate unlimited iPhone users unlimited data plans (July 2011) 9 (May 2012) 11 AT&T starts $10/GB T-Mobile starts Verizon AT&T introduce overage fee for data caps and shared data plans with smartphone data plans throttling penalty unlimited voice and text (June 2010) 6 (May 2011) 7 (June 2012) 12 Time-Warner trials AT&T caps U-verse to 250 GB data caps in and DSL to 150 GB with a $10 Texas penalty for an additional 50 GB (June 2008) 1 (May 2011) 4 Comcast moves towards tiered usage- AT&T starts throttling Comcast caps based billing wireline users data at 250 GB (May 2012) 5 (April 2011) 3 (August 2008) 2 Introduction of Data Caps Wireline

5 But Not Heavy All the Time How to leverage the peak-valley differential? Peak demand > 99% Average demand < 30%

6

7

8

9 Time Elasticity: Opportunities Opportunities Movies & Volume Multimedia Streaming downloads, videos, P2P Gaming Cloud Software Downloads Email, Texting, Social Weather, Network Finance updates Time Elasticity Pricing Survey: http://arxiv.org/abs/1201.4197

10

11

12

13

14

15

16

17

18

19 Contributions 1. An architecture and a fully functional system for offering TDP for mobile data 1. User behavior models and optimized price computation 1. A realistic evaluation with real users

20 TDP Overview Prices Internet Traffic

21 TUBE Theory

22 Feedback Loop Prices User Behavior Estimation User Network Interface Price Measurement Calculation Usage

23 Waiting Functions w • Probabilistically estimate “ willingness ” to wait r = impatience Estimate parameters r 1 m 1 w r 1 + m 2 w r 2 + m 3 w r 3 r 2 discount r 3 time waited

24 Minimizing Cost Exceeding capacity Traffic Offering discounts Time variables d i G 1 +G 2 minimize

25

26

27

28 TUBE Architecture

29 Design Guidelines 1. Separating functionality  Price computation on a central server  Price display and scheduling the usage on the user devices 2. Scaling up the system  User behavior estimation algorithm requires only aggregate, not individual usage data  Formulate the price calculation as a convex optimization for computation scalability for many TDP periods 3. Protecting user privacy  No Deep Packet Inspection (DPI) and no private data is exchanged. 4. Empowering user control

30 TUBE: TDP Architecture User Device ISP Server Secure Connection User GUI Price Usage Price Information Monitor optimizer Youtube User Autopilot Netflix Behavior ! "#$% & Flipboard Estimation Magazine ' (% ) * +, - . & / 00& Aggregate Apple 1+2% ) 34 % (& Traffic App Store Measurement Application Traffic Allow or Block

31 TUBE Implementation

32 Server Side Design User Device Queries REST API Delegation Pricing Policy Container Authentication Price Optimizer Mechanism Module Pricing Plans User Behavior Manual TDP Estimation TDP Autopilot Measurement Push Notifier Pricing Policy Enforcer Plugins Usage Monitoring Plugins Netfilter Message Standard Control RADIUS Standard Netfilter Traffic Linux Linux Linux SMS Push ! ! DPI 3G 3G

33 Server Side Design Number of TDP periods User Device Queries Number of Periods 12 24 48 96 144 REST API Behavior Estimation (sec) 12.76 200.0 959.6 1967 15040 Delegation Pricing Policy Container Authentication Price Calculation (sec) 1.67 1.69 1.70 1.81 1.84 Price Optimizer Mechanism Module Pricing Plans User Behavior Manual Number of TDP periods x Number of application types TDP Estimation TDP Number of Application Type Autopilot Measurement Number of Periods 2 4 8 12 0.21 12.99 21.52 Push Notifier Pricing Policy Enforcer Plugins Usage Monitoring Plugins 24 3.33 47.08 75.47 Netfilter Message Standard Control RADIUS Standard Netfilter Traffic Linux Linux Linux SMS Push ! ! DPI 3G 3G 48 15.99 197.22 215.42 (mins)

34 Client Side Design Graphical User Interface Top 5 Price Current Settings Usage Popup Price Apps Status bar Bill Display Display Display Display Display Display Display Daemon Enforcer Price Price DB Budget Communication Notifier Dispatcher Helper Module Budget Task Usage DB Manager Manager App Pulled Usage Scheduler Allow/Block Collector Local Delegation Autopilot Algorithm Usage Session Manual Budget App PPI Scheduler Tracker Recorder Autopilot Type Status bar App usage Daemon Support LOC iPhone No No Partial 25K Android Yes Yes Yes 5.4K Windows Yes Yes Yes 5.3K

35 TUBE Princeton Trial

36 Princeton Trial: Money Flow TDP based payments Current pricing scheme $$ $$ AT&T $$ Participants TUBE Project Wireless Provider • 50 AT&T participants: 27 iPhones, 23 iPads • Faculty, staff, and students • 14 Academic Departments & Divisions

37 Princeton Trial: Data Flow PSTN BSS 3G Core Network Gateway GMSC� MCS� AT&T Firewall DNS AuC� VLR� HLR� VPN Data BSC� Flow Data GGSN� Flow NAT User's SGSN� iPhone, AT&T's iPad mobile TUBE network Servers

38 TUBE App: Information Screens Price indicator 4 4 4 % % %

39 TUBE App: Scheduling Screens

40 Princeton Trial Results

41 Two main goals of the trial 1. How do people respond to pricing changes and GUI design? 1. Can the end to end TDP system work in the real world and can our architecture scale up?

42 Usage Statistics • How much bandwidth participants use? – „ Heavy tailed ‟ • Which applications use the most bandwidth – Streaming and surfing Movie� Type� Web� Non-Jailbroken� Applica on� Downloads� iPads� Music� Jailbroken� iPads� News/Mags.� iPhones� Other� 0� 0.1� 0.2� 0.3� 0.4� 0.5� Usage� Volume� (GB)�

43 UI Effectiveness Do users respond more to the numerical values of TDP prices or to the color of the price indicator bar on the home screen?  Users paid more attention to indicator color than the numerical discount values

44 UI Effectiveness  Users paid more attention to indicator color than the numerical discount values Type Periods First Stage Second Stage Color Discount Color Discount 1 2, 8, 14, 20 Orange 10% Orange 28% 3, 6, …, 24 2 Orange 10% Green 30% 3 5, 11, 17, 23 Orange 10% Orange 9% Period types 1 and 3 Period types 2 and 1

45 Optimized TDP Impact Does the peak usage decrease with time-dependent pricing? And does this decrease come at the expense of an overall decrease in usage? » Optimized TDP reduces the peak-to-average ratio » Overall usage significantly increases with TDP

46 Optimized TDP Impact Does the peak usage decrease with time-dependent pricing? And does this decrease come at the expense of an overall decrease in usage? » Optimized TDP reduces the peak-to-average ratio » Overall usage significantly increases with TDP Overall usage PAR reduces by 30% increases by 130%

47 Trial Limitations and Extensions Limitations: 1. Single bottleneck 2. Mobility 3. Control group 4. Time granularity Extensions: 1. Location/congestion dependent pricing 2. Commercial operator trials

48 Summary 1. A fully functional system for offering TDP for mobile data 2. People are sensitive to time-dependent prices and indeed shift their Internet usage to off-peak periods 3. The pilot trial motivates future study on TDP for different markets and demographics

49 TU BE Thank you! sangtaeh@princeton.edu  DataMi : http://www.datami.com  DataWiz : http://www.datawizapp.com  SDP Forum : http://scenic.princeton.edu/SDP2012

50 Backup Slides

51 TDP Performance

52 Price Sensitivity Do users wait to use mobile data in return for a monetary discount? » Average usage decrease in high-price periods relative to the changes in low-price periods

53 Notification Effectiveness Do notifications impact usage? » 80-90% of users decrease or did not increase their usage after the 1 st notification » For all subsequent notifications, about 60-80% of the active users decrease their usage, while the others remained price-insensitivity