1. Distribute what you can, centralize what you must!
Narseo Vallina-Rodriguez
Supervisor: Jon Crowcroft
Qualcomm – Cambridge 22nd May 2013
Apologize Jon; 2 day project meeting in London
Idea: current centralised model: cloud-phone does not really work

2. Motivation The web is becoming mobile
Apps rely on multiple online/cloud services (mobile mashup):
CDNs (Akamai)
Cloud services (Amazon WS)
Authentication APIs (Oauth)
Assisting sensors (A-GPS)
Advertisement (AdMob, Burstly, Millennial Media, …)
Push notifications (Google’s GCM)
NAT punching for P2P (Skype)

3. Research question
How do mobile apps’ cloud dependency impact on cellular network and battery life of mobile handsets?

4. outcome
When Assistance becomes Dependence: characterizing the costs and inefficiencies of A-GPS. Vallina-Rodriguez, Finamore, Grunenberger, Papagiannaki and Crowcroft. ACM SIGMOBILE MC2R (under review)
Breaking for Commercials: Characterizing Mobile Advertising. Vallina-Rodriguez, Finamore, Shah, Grunenberger, Haddadi, Papagiannaki and Crowcroft. In ACM Internet Measurement Conference 2012(IMC'12)
Energy Management Techniques in Modern Mobile Devices. Vallina-Rodriguez and Crowcroft. In IEEE Communications Tutorials and Surveys, 2012.
When David can help Goliath: the case for cellular augmentation of wired networks. Vallina-Rodriguez, Erramilli, Grunenberger, Gyarmati, Laoutaris, Stanojevic, Papagiannaki, In ACM HotNets'12
Signposts: End-to-End Networking in a World of Middleboxes. Aucinas, Chaudhry, Crowcroft, Probst Eide, Hand, Madhavapeddy, Moore, Mortier, Rotsos and Vallina-Rodriguez. In ACM SIGCOMM DEMO

5. Take away: moving to the edge!
Mobile applications may abuse cellular networks: they cause network (signaling/channels/operational) and energy costs!
Fetching content in a centralized fashion is not the only way
Apps and OS must exploit locality and neighboring devices when possible!

6. Distribute as much as you can!

7. Flashlinq/LTE-direct
P2P wireless technology
Perfect candidate for transparent communication in the edge!
Peer discovery (energy efficient)
Expression-based discovery (service)
Always-on background service with low duty-cycle
Similar to powering up a paging channel every X seconds
Current prototype performance:
Low-latency (<10 ms)
Good throughput (~ 20 Mbps)
Discovery (1~2 seconds)

8. … but what can be distributed?

9. 1. Localized data

10. Use case 1: Localized data
A large fraction of mobile data is local
Weather
Notifications
Ads
Apps use cellular networks and push notifications to fetch this content
High latency
No delivery guarantees
[Cellular data network infrastructure characterization and implication on mobile content placement, Xu et al. SIGMETRICS’2011]

11. Use case 1: Airport notifications
SERVER (UK)
NODE B
Google
GCM (Ir)
RNC
SGSN
GGSN
INTERNET

12. Use case 1: Airport notifications
Traffic Pattern Heathrow App For Android (Flight Update)
Energy Signaling Spectrum (HSPA)
TCP/IP Push notification model is broken for local data:
Frequent RNC promotions (some caused by TCP Heartbeats)
Waste of energy, middleboxes/proxies memory and radio channels (+200K users/day, a lot of signaling traffic!)

13. Use case 1: Airport notifications
PubSub model Low latency
No net overhead
Energy efficient
No Middleboxes!
SERVER (UK)
Flashlinq
NODE B
Google
GCM (Ir)
RNC
SGSN
GGSN
INTERNET

14. 2. Collaborative sensing

15. Use case 2: Collaborative A-GPS
Assisting data (time, ephemeris, almanac, coarse location) downloaded from network:
Reduces TTFF (usability)
Temporal validity up to 2 weeks for ephemeris
Problem: use of cellular network may impair performance and increase energy costs!

16. Use case 2: Collaborative A-GPS
2x current!
Control-plane latency

17. Use case 2: Collaborative A-GPS
Collaboration between devices in a P2P fashion:
Context-awareness (sense environment so do not turn on AGPS indoors!)
Share/pre-fetch assisting data (reduces latency to fetch data)
Prototype for Nexus One:
Pre-fetch and cache of assisting data
Devices can detect if they’re indoors in less than 10 seconds
Blackbox. Hard to inject assisting data on chipsets (A-GPS is controlled by binary/proprietary files/drivers  )

18. 3. Wired-wireless integration
Thinking beyond app-level use cases,
Flashlinq is a perfect candidate for wired-wireless integration.

19. Use case 3: Wired-wireless integration
3G offloading to WiFi and femtocells:
Reduce network traffic
No real benefit for users (unless volume cap in data-plan)
Wired network can be constrained!
Can cellular networks augment wired networks?
Wired nets deployment is $$$
Cellular nets have good coverage

20. Use case 3: Wired-wireless integration
Cellular network can provide more capacity than wired ones (DSL)
Spare capacity on cellular network
Powerboost for video- streaming apps
Use-and-release
Does NOT work everywhere anytime!
4.7 Mbps
DSLAM
DSL in rural/suburbs: far from DSLAM
Spare capacity for small periods of time (HSPA)
Powerboost (use and release) A
2 Km
2.8 Mbps
Google Maps

21. Use case 3: Wired-wireless integration
2x downlink/5x uplink for most locations with 1 mobile device
Simulation: 50% of the videos have a speed up factor of 10x
2x downlink
5x uplink
Some locations are congested
Nevertheless, worst case shows that 50% apps can reduce by 10x the buffering time

22. Conclusions Current cloud-mobile model is not efficient
Hyper-centralized: push notifications
Lack of connectivity between handsets: missing opportunities
Cellular and wired networks are fully decoupled
Flashlinq/LTE-direct can bring a new mobile paradigm!
Energy and network efficient
Distributed
Flexible

23. Flashlinq limitations and extensions
Transparent security/authentication mechanisms
Lessons to be learnt from the past: Bluetooth and WiFi-direct failed!
Source of DoS/Privacy/Energy attacks
Global Signpost-ish naming (OpenSource, DNSSEC based)
Low-level radio details must be exposed to OS!
Too much layering hides inefficiencies: e.g. A-GPS and 3G
Cross-layer optimizations are key (e.g. iPhone vs. Android)
Incentives for operators?
Reduce operational costs: better use of limited capacity
Licensed frequency
.. But there are few missing bits from my perspective as a developer that are being missing

24. Thank you for your attention!