220 likes | 519 Views
CAMEO: A Middleware for Mobile Advertisement Delivery. Azeem Khan† , Kasthuri Jayarajah *, Dongsu Han ‡, Archan Misra *, Rajesh Balan *, Srinivasan Seshan ‡ * Singapore Management University ‡ Carnegie Mellon University †Oriental Institute of Management. Motivations.
E N D
CAMEO: A Middleware for Mobile Advertisement Delivery Azeem Khan†, KasthuriJayarajah*, DongsuHan ‡, ArchanMisra*, Rajesh Balan*, SrinivasanSeshan‡ * Singapore Management University ‡ Carnegie Mellon University †Oriental Institute of Management
Motivations • Improving performance of mobile advertisements delivery • Decreasing bandwidth usage • Reducing energy consumption on mobile • Introduce monetization of advertisements by users and ISPs
Research Challenges • Reduce overheads in delivering ads • Provide offline access to advertisement selection • Framework that enables dynamic negotiation in trading advertisements for connectivity
Background for addressing performance issues 48% ads seen after 6hours • Data collection process for advertisements • Contexts: app, location, device type, OS, carriers • Period: Every 1 minute, 2 weeks – 2 months • Procedure: Scripts on computers in USA, Asia, Europe • Observations • Top 100 ads account for > 50% of views • 37% of ads are seen even after a day • 2/3 or more of ad content is redundant across ads (templatedHTML) • Country specific ads (overlap < 6%) 37% ads seen after a day
Background study for performance issues • Observations • On average, users switch between WiFi and 3G networks 2-4 times per day • Users are often connected to WiFi when charging phone • Users are on 3G network more than 50% of the time • Data collection procedure for users • Who: 20 participants on SMU campus for 1 month • Procedure: Custom LiveLabs app running as a monitoring service on Android 4.0+ WiFi connected Heavy WiFi usage
Challenge #1: Reduce overheads How? Pre-fetching and caching of advertisements. Why Both? • pre-fetching • CAMEO exploits the fact that users are often on cheaper WiFi networks more than once a day! • Advertisement contexts that matter such as location and app can be predicted • caching • Ads are repeated • Small number of ads account for most ad views • Overheads per ad are avoided
Caching and Pre-Fetching AN APP #1 APP #2 More than 70% savings in ad related bandwidth is observed… AN CONTEXT PREDICTOR AD MANAGER CACHE CAMEO AN = ADVERTISEMENT NETWORK
Challenge # 2: Offline Access to Ads • Online selection of ads • AN advertisement selection (ANAS) • Bulk pre-fetch of ads, online ad selection by AN • Offline selection of ads • Local advertisement selection (LAS) • Bulk pre-fetch of ads, AN provides selection rules • Best effort advertisement selection (BEAS) • Bulk pre-fetch of ads, statistical selection by CAMEO
Advertisement Selection APP #1 APP #2 AD MANAGER ACCOUNTING & VERIFICATION RULESET CACHE CAMEO AN = ADVERTISEMENT NETWORK
Energy gains by pre-fetching and caching • Base case measurement procedure • Screen is lit (50% brightness on Samsung S3) • No other app/services running except OS default • WiFi of SMU campus, 3G on SingTel Singapore • Pre-fetching performed on cheaper network when phone is charging. • ads fetched once every 45 seconds by custom app • Monsoon monitoring device measures device power consumption • Gains in LAS and BEAS for 1000 ad views for mostly offline apps • 99% savings in energy of radio use And nearly 92% savings in bandwidth
Challenge # 3: Bartering ads for connectivity • Example Scenario: A man walks into a airport where they charge $10 for connectivity. Would it be possible for him to get access in exchange for seeing advertisements from the airport’s network? • Implications & Assumptions • Foreground apps • Negotiations are transparent to the user
Can we trade? 1. BARTER? APP 4. AD(S) ISP CAMEO 2. NEGOTIATE 5. BITS USED 3. AD FETCH OS
CAMEO architecture APP #1 APP #2 ISP NEGOTIATOR AD MANAGER AN# 1 LIBRARY CONTEXT PREDICTOR AN #2 LIBRARY ACCOUNTING AND VERIFICATION CAMEO
Limitations of current CAMEO implementation • The user study is not representative • Long term context prediction may never be 100% accurate • A small amount of space in memory will be occupied by the cache (approx. 2 MB for 1000 ad views) • Accounting and verification need to be robust. These issues are currently under investigation.
Summary • #Challenge 1: Reduce overheads • Pre-fetching and caching enable significant reduction in bandwidth and energy consumption • #Challenge 2: Offline access of ads • online and offline modes of ad selection to preserve and enhance current economic models • #Challenge 3: Framework for trading • Initial framework proposed and implemented *Thanks to Matt Welsh, the PC reviewers and my colleagues at SMU*
Mobile Advertising Stakeholders AN ≡ advertising network Signaling overhead Bandwidth Quota Energy consumption
EMPIRICAL STUDY - Advertisements Caching could be very effective Large amounts of redundant information Small percentage of ads dominate views
EMPIRICAL STUDY - Users Users are mostly on expensive networks Users are price conscious
Design Goals • Lower cost of advertisement delivery • Minimize user involvement • Incentivize developers to make applications consumer friendly • Minimal modifications to applications and mobile advertising networks.
CONTEXT PREDICTION Algorithm to analyze and predict context Context prediction accuracy
CAN WE TRADE? ISP NEGOTIATOR ISP G/W APP CAMEO ANDROID OS Register (1 ad, 10KB, TCP port 2894) Negotiate Accepted Success Request Ad Context Specific Ad Ad ready 10 KB, TCP 2894 IP A.B.C.D Display Ad How many bytes? Data transmission Accounting Accounting Disconnect Thanks for all the fish Close 2984 Bye Completed