Edge10 Workshop on Princeton Edge Lab’s 10th Mung Chiang May 17, 2019

Outline Ten years ago… Smart Data Pricing Edge/Fog Edge for Pricing Pricing the Edge Edge/Fog SCALE Interfaces Fogonomics Dispersive Learning

Acknowledgments Postdocs, students, visitors Collaborators Funding agencies Industry partners

0. Ten Years Ago…

Research Bridget theory-practice gaps in networking Proofs to prototypes Edge/Fog (technological networks) Smart Data Pricing (economic networks) Social Learning Networks (social networks)

Education 2011: Network20Q & flip classroom 2012: MOOC (Chris) 400,000 students 2012: “Networked Life” 2016: “Power of Networks” (Chris)

Startups 2013: DataMi (Sangtae, Carlee, Soumya) 60 million users 2014: Zoomi (Sangtae, Ruediger, Chris) 2015: Smartiply (Junshan, Kaushik) 2017: Myota (Jaeyoon, Sangtae)

Industry and Community Impact About a dozen company partners 2015: OpenFog Consortium 2018: Industrial Internet Consortium Major conference panels, workshops, industry forums Special journal/magazine issues and 2 edited books on Fog & SDP ~50 postdocs/Ph.D. students, ~25 as faculty and ~25 in industry

1. Smart Data Pricing

SDP Dimensions How? Whom? What? More Usage-based, demand response … real-time … Whom? Toll-free (1-800, zero rating, sponsored data, split billing)… What? App-based (no data plan), cloud pricing… IoT pricing, PMP… More Offloading, Quota-aware preloading…B2B, roaming, peering… AT&T speed tiers

Example: Time Elasticity of Applications Large Peak-Valley Differential Streaming videos, Gaming Texting, Weather, Finance Email, Social Network updates Cloud Software Downloads Movies & Multimedia downloads, P2P Opportunities Opportunities for Exploiting time-elasticity of demand

Cost-effective Mobile Content Delivery Reduce peak & increase valley Defer capital spending Sell unused capacity Increase revenue

Edge Complements Cloud SDK

Rethink Spectrum Flashy Whitespace

Rethink Ecosystem Stop (just) counting bytes and start living with QoE Recognize, leverage heterogeneity of apps and networks Win – Win – Win Consumers: more choices and lower $/GB Carriers: higher revenue and lower cost Content and app providers: more engaged eyeballs

Rethink Networks End User Cellular Core Smart sharing in APP + PHY Mobile management from the edge

Pricing 5G Spectrum allocation/auctions for new bands of licensed and unlicensed spectrum Infrastructure sharing:  given densification, how will resource sharing work between competitive operators? Pricing of consumer mobile Pricing for broadband access Pricing of industrial IoT How will these pricing options evolve when killer apps emerge and mmWave devices become affordable? Taken from JSAC special issue proposal.

Pricing IoT How to charge? Whom to charge? Time-dependent? Volume discounts? Application-dependent pricing: pricing with guarantees on delivered outcome or experience, e.g. price 5G network slices with guaranteed QoS Whom to charge? Stakeholders include IoT service provider (e.g. smart home sensors), IoT wireless access provider (e.g AT&T), and IoT cloud platforms (e.g. Amazon AWS IoT Hub) Whom should users pay? Users pay each separately vs users pay only service provider vs … Vertical integration of stakeholders What happens if AT&T or Verizon offer both IoT management platforms and connectivity? (they do; Verizon ThingSpace and AT&T Control Center)

2. Edge/Fog

Distribute functions to network edge 2009 Distribute functions to network edge

Distribute functions along Cloud-2-Things Continuum 2015 2018 Distribute functions along Cloud-2-Things Continuum

To Fog or Not to Fog: SCALE Security Cognition Agility Latency Efficiency

Fog as An Architecture Architecture is “Horizontal Foundation”: Who does what, at what timescale, how to glue them together? Allocation of functions, not just resources Architecture supports Applications: Source-channel separation: Digital communication TCP/IP: Internet applications Fog/edge: IoT / 5G / Dispersive AI

A. Interfaces Massive storage Real time processing Heavy duty computation Global coordination Wide-area connectivity Real time processing Rapid innovation Client-centric Edge resource pooling

Example: Shred and Spread Client-driven data processing for privacy protection and reliability Scatter files to multiple fog storages Client-side data deduplication Obfuscated data in storages File chunking for data deduplication Chunk encoding/spreading for privacy and reliability

Example: Networked Drone Cameras

S S B. Fogonomics Compute price: Memory size Compute time Data storage Communication price: Requests across functions Data transmission Internet access Incentivizing local dispersed resources: Cellular data plans User mobility pattern Heterogeneous devices Network connections

Application-Dependent Pricing The specific application offered changes resources and pricing Example: Pricing of data collected by edge devices Optimal amount and frequency of charging Pricing based on measures of freshness of data How to price and sell private data? Example: Pricing of distributed ML services Using fog/edge resources for distributed ML to make inferences, find correlations, or for online planning

C. Edge/Fog for Dispersive AI Design machine learning algorithms that support fast responses Decompose machine learning into multiple geographically distributed components (jointly operating to adaptively optimize data collection/analytics) Minimize communication costs and centralized data processing costs Make best use of local/proximal resources Proactively pre-position content and computing Parallelize successive refinement for streaming mining Reduce infrastructure costs and improve quality of experience

Dispersive Learning Decentralized, online decision making under uncertainty by a team of edge devices in an unknown environment Examples: fleet of drones deployed for anti-poaching efforts, team of disaster relief robots Solution approach: multi-agent reinforcement learning, augmented with inter- agent communication for better learning and coordination Information shared by the informed devices with others could in fact degrade their learning early on Delayed sharing may be preferred: wait until policies have improved, then share

Information sharing might help learning… Timing Matters Information sharing might help learning… Or might degrade it! P. Naghizadeh, M. Gorlatova, A. Lan, M. Chiang. “Hurts to Be Too Early: Benefits and Drawbacks of Communication in Multi-Agent Learning”, INFOCOM 2019.

Unique Challenges & Opportunities Heterogeneity/Under-organization of resources/devices Variability/Volatility in availability/mobility Constraints in bandwidth/battery Proximity to sensors/actuators

Thank you & To the next 10 years chiang@purdue. edu chiangm@princeton Thank you & To the next 10 years chiang@purdue.edu chiangm@princeton.edu