Traffic Offloading and Wireless Edge Networks: Theory and Novel - - PowerPoint PPT Presentation

Traffic Offloading and Wireless Edge Networks: Theory and Novel Realizations

Leandros Tassiulas Yale University WiOpt, Paris, May 2017.

A New Era in Wireless Networking

• Recent developments:
• Mobile data traffic growth, new services & advanced devices, 5G vision.
• Challenges for cellular networks:
• Accommodate the growing traffic.
• Support emerging 5G services.
• Traditional network expansion methods:
• Upgrading technology, acquiring new spectrum, deploying more cells, ...

... are costly and time-consuming solutions.

• Our approach:
• Explore methods that aim to fully utilize (i) existing spectrum allocations and

(ii) idle user-owned wireless infrastructure.

Outline

• Mobile data offloading.
• Use Wi-Fi capacity to serve cellular traffic.
• User Provided Networks (UPNs).
• Facilitate multi hop wireless access through exchange of wireless resources.
• Prototype realization based on mobile SDN.
• Resource exchange markets in networks.
• An Arrow-Debreu type formulation for networks.

Mobile Data Offloading

Offloading can be realized over femtocells or Wi-Fi access points.

• Goal: reduce network costs (OpEx) & accommodate more traffic.
• The wireless spectrum or the wired link are not owned by the operator.

Potential and Key Question

• Mobile network operators have adopted such solutions:
• AT&T had deployed 32,000 Wi-Fi hotspots by 2012.
• T-Mobile and other operators, collaborate with FON.
• Republic Wireless, Google Fi, etc.
• Offloading benefits depend on the availability of APs’.
• How to increase this availability?
• Our proposal for operators: lease idle user-owned Wi-Fi APs.
• Residential Wi-Fi APs are often underutilized.
• On-demand & low-cost network capacity expansion.
• Fully aligned with 5G design principles.

Mobile Data Offloading Markets

• Related Publications:
• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, An Iterative Double Auction for

Mobile Data Offloading, IEEE WiOpt, 2013, (Best paper Award), IEEE/ACM

• Trans. on Networking, 23(5), 2015.
• K. Poularakis, G. Iosifidis, L. Tassiulas, Deploying Carrier-grade WiFi:

Offload Traffic, Not Money, ACM Mobihoc, 2016.

• L. Gao, G. Iosifidis, J. Huang, L. Tassiulas, D. Li, Bargaining-Based Mobile

Data Offloading, IEEE JSAC, SI on 5G, 32(6), 2014.

• A. Apostolaras, G. Iosifidis, K. Chounos, T. Korakis, L. Tassiulas, A

Mechanism for Mobile Data Offloading to Wireless Mesh Networks, IEEE

• Trans. on Wireless Comm., 15(9), 2016.
• K. Poularakis, G. Iosifidis, I. Pefkianakis, L. Tassiulas, Mobile Data

Offloading through Caching in Residential 802.11 Wireless Networks, IEEE

• Trans. on Network Services & Management, 13(1), 2016.

Data Offloading Marketplace

• A set of network operators:
• Each operator owns many base stations.
• Each BS had different load.
• A set of access points:
• Each AP has different Internet capacity.
• AP owners have communication needs.
• A Broker
• Goal & Key questions:
• Efficiency: maximize BSs’ benefits, minimize APs’ costs.
• How much traffic from each BS should be offloaded to each AP?
• How much each AP owner should be reimbursed for serving this traffic?
• Technical Issues:
• Offloading Benefits are AP-specific and interdependent.
• Offloading Capacities of the APs are coupled.

Data Offloading Marketplace

• A set of network operators:
• Each operator owns many base stations.
• Each BS serves different amount of traffic.
• A set of access points:
• Each AP has different Internet capacity.
• AP owners have communication needs.
• A Broker
• Economic Issues:
• Multiple buyers & multiple sellers with conflicting goals.
• Information asymmetry about the needs.
• Solution approach:
• Use an auction to elicit hidden information.
• Traditional auctions, e.g., VCG and McAfee, cannot be used.
• Design a new auction algorithm.
• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, A Double Auction Mechanism for Mobile Data

Offloading Markets, IEEE/ACM Trans. on Networking, vol. 23, no. 5, 2015.

Data Offloading Marketplace

• A set of network operators:
• Each operator owns many base stations.
• Each BS serves different amount of traffic.
• A set of access points:
• Each AP has different Internet capacity.
• AP owners have communication needs.
• A Broker
• Economic Issues:
• Multiple buyers & multiple sellers with conflicting goals.
• Information asymmetry about the needs.
• Solution approach:
• Use an auction to elicit hidden information.
• Traditional auctions, e.g., VCG and McAfee, cannot be used.
• Design a new auction algorithm.
• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, A Double Auction Mechanism for Mobile Data

Offloading Markets, IEEE/ACM Trans. on Networking, vol. 23, no. 5, 2015.

Data Offloading Marketplace

• A set of network operators:
• Each operator owns many base stations.
• Each BS serves different amount of traffic.
• A set of access points:
• Each AP has different Internet capacity.
• AP owners have communication needs.
• A Broker
• Economic Issues:
• Multiple buyers & multiple sellers with conflicting goals.
• Information asymmetry about the needs.
• Solution approach:
• Use an auction to elicit hidden information.
• Traditional auctions, e.g., VCG and McAfee, cannot be used.
• Design a new auction algorithm.
• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, A Double Auction Mechanism for Mobile Data

Offloading Markets, IEEE/ACM Trans. on Networking, vol. 23, no. 5, 2015.

Model

• A market of multiple BSs and multiple APs, studied for a period T:
• M, {1, ..., M}: the set of BSs; I, {1, ..., I}: the set of involved APs.
• Base station m:
• xm, {xm1, ..., xmI}: offload request vector.
• Jm(xm): offloading benefit.
• Access Point i:
• Ci: Internet access capacity.
• yi, {yi1, ..., yiM}: offload admission vector.
• Vi(yi): offloading cost.
• Broker’s objective: Efficiency Maximization

maximize

xm,yi ,8m,8i

X

m2M

Jm(xm) X

i2I

Vi(yi) Efficiency

subject to (i)

P

m2M yim  Ci, 8i 2 I,

Capacity constraint

(iii) xmi = yim, 8m 2 M, i 2 I.

Feasibility

Iterative Double Auction – IDA

• An auction mechanism includes:
• An allocation rule & a pricing rule.
• Bidders’ Bidding Problems
• BS Bids: pm = (pmi : i = 1, . . . , I) .

Pm : maximize

pmi 0,8i

Jm(xm(pm)) hm(pm), for every BS m;

• AP Bids: αi = (αim : m = 1, . . . , M) .

Pi : maximize

αim0,8m

Vi(yi(αi)) + li(αi), for every AP i.

• Broker’s Allocation Problem

maximize

xm,yi ,8m,8i

X

m2M

X

i2I

⇣ pmi log xmi αim 2 y 2

im

⌘

subject to (i)

P

m2M yim  Ci, 8i 2 I, (ii) xmi = yim, 8m 2 M, i 2 I.

Iterative Double Auction – IDA

• The KKT conditions for the efficiency maximization problem:

(A1) : ∂Jm(x

m)

∂xmi = µ

mi, (A2) : ∂Vi(y i )

∂yim = µ

mi λ i ,

(A3) : λ

i ·

⇣

M

X

m=1

y

im Ci

⌘ = 0, (A4) : x

mi = y im,

(A5) : µ

mi · (y im x mi) = 0, (A6) : x mi, y im, λ i 0.

• The KKT conditions for the broker problem:

(B1) : x⇤

mi = pmi

µ⇤

mi

, (B2) : y ⇤

im = µ⇤ mi λ⇤ i

αim , (B3) (B6) = (A3) (A6)

• If APs and BSs submit:

pmi = x⇤

mi · ∂Jm(x⇤ m)

∂xmi , αim = 1 y ⇤

im

· ∂Vi(y⇤

i )

∂yim ... the solutions coincide.

Iterative Double Auction – IDA

• The KKT conditions for the efficiency maximization problem:

(A1) : ∂Jm(x

m)

∂xmi = µ

mi, (A2) : ∂Vi(y i )

∂yim = µ

mi λ i ,

(A3) : λ

i ·

⇣

M

X

m=1

y

im Ci

⌘ = 0, (A4) : x

mi = y im,

(A5) : µ

mi · (y im x mi) = 0, (A6) : x mi, y im, λ i 0.

• The KKT conditions for the broker problem:

(B1) : x⇤

mi = pmi

µ⇤

mi

, (B2) : y ⇤

im = µ⇤ mi λ⇤ i

αim , (B3) (B6) = (A3) (A6)

• If APs and BSs submit:

pmi = x⇤

mi · ∂Jm(x⇤ m)

∂xmi , αim = 1 y ⇤

im

· ∂Vi(y⇤

i )

∂yim ... the solutions coincide.

Iterative Double Auction – IDA

• The payment and reimbursement rules we employ are:

hm(pm) =

I

X

i=1

pmi, m = 1, . . . , M li(αi) =

M

X

m=1

yim(λi µmi), i = 1, . . . , I

− −

20 40 60 80 100 120 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0.2 0.4

Gap y − x Step − t

BS 1, AP 1: y11 −x11 BS 1, AP 2: y21−x12 BS 2, AP 1: y12−x21 BS 2, AP 2: y22−x22

Iterative Double Auction – IDA

BS2 AP1 BS1 AP2 AP3

BROKER

Broker announces pricing signals (Lagrange Multipliers) 1

1 The broker announces the pricing signals

λi, µmi, i 2 I, m 2 M.

Iterative Double Auction – IDA

BS2 AP1 BS1 AP2 AP3

BROKER

Each AP find its currently optimal bid vector 2 Each BS finds its currently optimal bid vector 2

1 The broker announces the pricing signals

λi, µmi, i 2 I, m 2 M.

2 Each AP i and BS m updates its bids

using the new Lagrange multipliers.

Iterative Double Auction – IDA

BS2 AP1 BS1 AP2 AP3

BROKER

Each BS sends its bids to the broker 3 Each AP sends its bids to the broker 3

1 The broker announces the pricing signals

λi, µmi, i 2 I, m 2 M.

2 Each AP i and BS m updates its bids

using the new Lagrange multipliers.

3 APs and BSs send their bids to the

broker.

Iterative Double Auction – IDA

BS AP1 BS AP2 AP3

BROKER

The Broker updates the Lagrange Multipliers 4

1 The broker announces the pricing signals

λi, µmi, i 2 I, m 2 M.

2 Each AP i and BS m updates its bids

using the new Lagrange multipliers.

3 APs and BSs send their bids to the

broker.

4 The broker updates the pricing signals λi,

µmi, i 2 I, m 2 M, using a gradient update.

Iterative Double Auction – IDA

BS AP1 BS AP2 AP3

BROKER

The Broker updates the Lagrange Multipliers 4

1 The broker announces the pricing signals

λi, µmi, i 2 I, m 2 M.

2 Each AP i and BS m updates its bids

using the new Lagrange multipliers.

3 APs and BSs send their bids to the

broker.

4 The broker updates the pricing signals λi,

i 2 I, µm, m 2 M, using a gradient update method.

5 The above steps are executed iteratively

until convergence: demands match

• fferings.

Summary

• A mobile data offloading market for leasing idle capacity.
• A new auction algorithm that achieves the optimal solution.
• No need to know offloading needs and cost functions.
• Can use a detailed network modeling approach.
• Other market models are also important to explore.
• Example: MNO reimburses its subscribers to open their APs.
• Caching-at-the-edge solutions employing user-owned APs.
• Intel’s 2011 experiments showed 40% reduction in backhaul traffic.
• L. Gao, G. Iosifidis, J. Huang, L. Tassiulas, D. Li, Bargaining-based Mobile Data Offloading,

IEEE JSAC, SI on 5G, 32(6), 2014.

• K. Poularakis, G. Iosifidis, et al, Mobile Data Offloading through Caching in Residential 802.11

Wireless Networks, IEEE Trans. on Network Services & Management, 2016.

User Provided Networks (UPNs)

• Related publications:
• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, Enabling Crowdsourced Mobile Internet

Access, IEEE Infocom, 2014, cond. accepted in IEEE/ACM ToN 2016.

• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, Incentive Mechanisms for User-provided

Networks, IEEE Communications Magazine, 52 (9), Sep., 2014.

• L. Gao, G. Iosifidis, J. Huang, L. Tassiulas, Hybrid Data Pricing for Network-Assisted

User-Provided Connectivity, IEEE Infocom, 2014.

• D. Syrivelis, G. Iosifidis, D. Delimbasis, K. Chounos, T. Korakis, L. Tassiulas, Bits and

Coins: Supporting Collaborative Consumption of Mobile Internet, IEEE Infocom, 2015.

• D. Giatsios, G. Iosifidis, and L. Tassiulas, Mobile edge-Networking Architectures and

Control Policies for 5G Communication Systems, WiOpt, 2016.

User Provided Networks

WiFi/Bluetooth Malfunctioned Congested WiFi Cell High-Performance WiFi/Bluetooth Cell

• Indicative Applications:
• Provide infrastructure connectivity to devices with no access.
• Overcome poor coverage through smart relaying.
• Support throughput-hungry services.
• Alleviate congestion problems or temporal malfunctions of the infrastructure.
• Innovative startups have already presented such solutions.
• Question: Can we find an optimal operation policy?
• How to aggregate and share the users’ network resources in an efficient and

fair fashion, such that users have enough incentives to participate?

• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, Enabling Crowdsourced Mobile Internet Access,

IEEE Infocom, 2014, cond. accepted in IEEE/ACM ToN.

User Provided Networks

WiFi/Bluetooth Malfunctioned Congested WiFi Cell High-Performance WiFi/Bluetooth Cell

• Indicative Applications:
• Provide infrastructure connectivity to devices with no access.
• Overcome poor coverage through smart relaying.
• Support throughput-hungry services.
• Alleviate congestion problems or temporal malfunctions of the infrastructure.
• Innovative startups have already presented such solutions.
• Question: Can we find an optimal operation policy?
• How to aggregate and share the users’ network resources in an efficient and

fair fashion, such that users have enough incentives to participate?

• G. Iosifidis, L. Gao, J. Huang, L. Tassiulas, Enabling Crowdsourced Mobile Internet Access,

IEEE Infocom, 2014, cond. accepted in IEEE/ACM ToN.

Proposed Solution

• A mechanism based on the Nash bargraining solution + virtual currency.
• Users are modeled through payoff functions.
• Utility from consuming data, energy cost and monetary cost for serving data,

virtual currency benefits.

• Efficiency and Fairness are addressed by the Nash Bargaining Solution.
• Pareto optimal.
• Takes into account the standalone operation of each node.
• Self-enforcing, hence users agree to apply the policy.
• Virtual Currency solves the double coincidence of needs and wants

problem.

• Decentralized implementation is possible if necessary.
• Dual decomposition of a convex optimization problem (the NBS problem).

Model

• A directed network G = (N, E) that we study for a period T.
• In(i): parent nodes of i, Out(i): child nodes of i.
• Each node n ∈ N can initiate a data session (n).
• Cij: capacity of (i, j) ∈ E , C0i: Internet capacity of i ∈ N.
• x(n)

ij

: bytes transferred over link (i, j), for commodity (n).

• y(n)

i

: bytes downloaded by node i, for commodity (n).

• Ui(ri): utility function modeling his communication needs, where

ri = y(i)

i

+ X

j2In(i)

x(i)

ji

• Vi(ei): energy consumption aversion function, where ei is the total

consumed energy in T: ei = X

j2Out(i)

es

ij

X

n2N

x(n)

ij

+ X

j2In(i)

er

ji

X

n2N

x(n)

ji

+ e0i X

n2N

y(n)

i

.

• pi ≥ 0: Internet access cost per byte.
• The overall payoff Ji(·) is defined as:

Ji(x, y) = Ui(ri) − Vi(ei) − pi X

n2N

y(n)

i

Problem Formulation

• Standalone performance: no relaying to/from others.

max

y(i)

i

C0i

Ji(y (i)

i )

Benchmark (minimum) performance Js

i = Ji(y ⇤ (i) i

).

• Virtual Coins system:
• Di: the initial coin budget of each user i ∈ N.
• Hi(·): the coins valuation function of user i (linear).
• z(n)

ij

: coins paid by j to i, for commodity (n), with (i, j) ∈ E.

• Each user receives γ > 0 coins for his participation in each round.
• Total budget of K coins in the system (upper bound).
• The payoff of each user includes now the coin budget:

JG

i (x, y, z) = Ji(x, y) + Hi(z, Di + γ)

Bargaining Problem

• The bargaining equilibrium can be derived by the solution of the

following convex problem. max

x,y,z

X

i2N

log

• JG

i (x, x, y) Js i Hi(z, Di)

• s.t.

X

j2In(i)

x(n)

ji

+ y (n)

i

= X

j2Out(i)

x(n)

ij , 8 i, n 2 N, i 6= n,

(1) X

n2N

x(n)

ij

 Cij, 8 (i, j) 2 E, X

n2N

y (n)

i

 C0i, 8 i 2 N (2) X

n2N

X

j2In(i)

z(n)

ji

• X

n2N

X

j2Out(i)

z(n)

ij

 Di + γ, 8i 2 N (3) JG

i (x, x, y) Js i + Hi(z, Di), 8 i 2 N

(4) x(n)

ij

0, y (n)

i

0, 0  z(n)

ij

 K, 8 i, j, n 2 N (5)

• where eq. (4) ensures that each user will receive a payoff at least equal

to his standalone performance.

User Provided Networks

• Question: Can we implement such systems in practice?
• Requirements:
• Independent of the physical layer.
• Highly adaptive to changing network conditions and users’ needs.

User Provided Networks

• Question: Can we implement such systems in practice?
• Requirements:
• Independent of the physical layer.
• Highly adaptive to changing network conditions and users’ needs.

CoNeS: Collaborative Network Sharing System

• Basic components:
• SDN-enhanced mobile devices: implement a programmable packet

forwarding datapath.

• Cloud service: monitors the nodes, and devises the policy.
• D. Syrivelis, G. Iosifidis, D. Delimpasis, K. Chounos, T. Korakis, L. Tassiulas, Bits & Coins:

Supporting Collaborative Consumption of Mobile Internet, IEEE Infocom, 2015.

CoNeS: Collaborative Network Sharing System

Internet

ISCD Service CDE SMDP Service Network Data Collection Decision Graph Derivation 3G/4G WiFi D2D links D2D data exchange Downloading / Uploading Data Statistics, Demands, Resources, Discovery Decision Graph

Gateway Relay/Client Client

Device Characteristics · Internet access capacity · Internet access cost · D2D links capacity · Battery energy Decision Graph D2D links Statistics & Demands SDN Control Plane SDN Data Path

Client Client

MBaaS Platform

2 1 1

• 1: Every node executes neighbor discovery.
• 2: Forwards to the cloud the network information (D2D links capacity), its

resource availability (battery, Internet throughput), and its demand.

CoNeS: Collaborative Network Sharing System

Internet

ISCD Service CDE SMDP Service Network Data Collection Decision Graph Derivation 3G/4G WiFi D2D links D2D data exchange Downloading / Uploading Data Statistics, Demands, Resources, Discovery Decision Graph

Gateway Relay/Client Client

Device Characteristics · Internet access capacity · Internet access cost · D2D links capacity · Battery energy Decision Graph D2D links Statistics & Demands SDN Control Plane SDN Data Path

Client Client

MBaaS Platform

3 3 4 4

• 3: The CDE collects the information; derives the servicing policy.
• 4: The decision graph is communicated to the nodes of the swarm.

Steps 1 - 4 are executed periodically.

CoNeS: Collaborative Network Sharing System

Internet

ISCD Service CDE SMDP Service Network Data Collection Decision Graph Derivation 3G/4G WiFi D2D links D2D data exchange Downloading / Uploading Data Statistics, Demands, Resources, Discovery Decision Graph

Gateway Relay/Client Client

Device Characteristics · Internet access capacity · Internet access cost · D2D links capacity · Battery energy Decision Graph D2D links Statistics & Demands SDN Control Plane SDN Data Path

Client Client

MBaaS Platform

3 3 4 4

• 3: The CDE collects the information; derives the servicing policy.
• 4: The decision graph is communicated to the nodes of the swarm.

Steps 1 - 4 are executed periodically.

Inside the Node

• OVS - Switch

HTB1 Queues

Port 1

Bluetooth Phy HTB2 Queues

Port 2

WiFi Phy HTBN Queues

Port N

LTE Phy

Local Port Virtual Ethernet Local IP Stack Mobile Node SMD (Linux Kernel) ICSD cfs dcs VPN Default Internet Gateway Tunnel OpenFlow API

• Open vSwitch datapath:
• Remotely configured, controls all network interfaces.
• Internet Connection Sharing Daemon (ICSD):
• Runs a discovery protocol & reports to CDE; gets & applies updates.

Performance Evaluation

• How often should the devices send status to the cloud?
• How fast is it possible to reconfigure the network?
• How much is the delay, bandwidth and energy consumption overhead?

Performance Evaluation

• How often should the devices send status to the cloud?
• How fast is it possible to reconfigure the network?
• How much is the delay, bandwidth and energy consumption overhead?

Performance Evaluation

• How often should the devices send status to the cloud?
• How fast is it possible to reconfigure the network?
• How much is the delay, bandwidth and energy consumption overhead?

Experimental Setup

• Embedded Nodes (single-board computers):
• Intel Atom CPU, 1Gbyte RAM,
• 802.11n WiFi (ad hoc mode), 100Mbit cable Ethernet interface.
• Real-time power consumption measurement.
• The cloud service is deployed at the NITOS cluster.

Experiments Findings

• 1

2 3 Internet Internet Gateway Gateway Client (1) (2)

• Status updates:
• 3 sec is optimal, 2.5% energy consumption, no additional delay. More

frequent updates double energy consumption.

• SDN Overheads:
• No important energy consumption or computation overheads (2%).
• Network reconfiguration:
• Gateway switching every 20 sec increases delay 24%, and energy

consumption 15%

Service & Resource Exchange over Networks

• Basic features of the system:
• Each node has some amount of spare resource.
• Nodes are complementary in terms of resource types or resource

availability.

• Their cooperation is constrained by a graph.
• Unsaturated demand.
• Indifferent in neighbors’ resources.

Service & Resource Exchange over Networks

• Various decentralized technological networks:
• Peer-to-peer file sharing overlays.
• Wireless mesh networks, Wi-Fi communities, Mobile Internet sharing.
• Renewable energy sharing in smart grid.
• Sharing economy platforms:
• Online bartering: swap.com, neighborgoods.net, etc.
• Food sharing, favor exchanging, risk sharing, etc.
• More examples: http://www.collaborativeconsumption.com/
• L. Georgiadis, G. Iosifidis, L. Tassiulas, ”Exchange of Services in Networks: Competition,

Cooperation, and Fairness”, ACM Sigmetrics, 2015.

Model

• An undirected connected graph G = (N, E).
• Set of allocations:

D =

• d = (dij)(i,j)2E : dij 0,

X

j2Ni

dij = Di}

• Set of feasible received resource vectors:

R =

• r = (ri)i2N : ri =

X

j2Ni

dji, i 2 N, d 2 D ,

• Individual node’s objective: to maximize his received resource ri, i 2 N.
• Exchange ratio vector:

ρi = ri Di , ρ = (ρi = ri Di : i 2 N)

Central Coordination Fair Allocations

• Question: Which is a sensible allocation?
• Ideal allocation: ri = Di, ∀ i ∈ N, i.e., ρi = 1
• Else: balance the exchange ratios as much as possible.
• Lexicographically optimal (Max-min fair) vector of exchange ratios ρ.
• There is a unique lex-optimal vector of exchange ratios ρ⇤ ⌫ ρ.
• Set R of received resource vectors is compact and convex, and ρi = ri/Di.
• Also interested in the allocations d⇤ that yield ρ⇤.
• While ρ⇤ is unique, there are many allocations d⇤ .
• Question: What are the main properties of ρ⇤ .

Properties of ρ∗

• There is a unique ρ⇤ and one or more d⇤ 2 D, with properties:
• Nodes are partitioned in distinct exchange ratio sets

L1, L2, . . . , L7.

• K = 7 depends on G and {Di} .
• L7 nodes work only with L1 nodes, and so on.
• It holds: l1 · l7 = l2 · l6 = . . . = 1.
• Topology: Lk is independent in the induced graph

GQk = (Qk, EQk ), k = 1, . . . , 3, where Qk = N − ∪k1

m=1(Lm ∪ LKm+1).

• Topology: L⇤

Kk+1 = NQk

• L⇤

k

• , k = 1, ...., 3.
• Theorem: If an allocation policy satisfies the above properties, then it is

lex-optimal.

Stability wrt Trade

• A Competitive Market.
• Every node i ∈ N determines independently his allocation policy
• dij
• j2Ni
• Objective: maximize P

j dji, or, equivalently, the ratio ρi = ri/Di.

• Ratio ρi can be interpreted as the price that node i sells his resource.
• An allocation d⇤ is an exchange equilibrium iff 8i 2 N:
• dji = dij · ρi, ∀ j ∈ Ni .
• if dji > 0 for some j ∈ Ni, then ρj = mink2Ni ρk .
• Does an exchange equilibrium exist?
• General Equilibrium theory: equilibrium exists under some mild conditions.
• Existence conditions do not apply in the proposed model:
• Not all nodes are endowed with non-zero quantities.
• Prices are not given exogenously; instead, they are indirectly determined by

the nodes’ decisions.

• K. J. Arrow, G. Debreu, ”Existence of and Equilibrium for a Competitive Economy”,

Econometrica 22(3), 1954.

Stability wrt Trade

Theorem.

1 There is a lex-optimal allocation d⇤ under which every node i 2 N gives

resource to its neighbors in proportion to what it gets from them, i.e., d⇤

ji

d⇤

ij

= r ⇤

i

Di = ρ⇤

i , 8 j 2 Di . 2 The neighbors not receiving resource from i have higher ratio ρj, i.e.,

ρ⇤

j 1

ρ⇤

i

, 8 j 2 Ni Di .

3 If the allocation satisfies the above conditions, then it is lex-optimal.

• Interpretation:
• There is a lex-optimal allocation where every node i ∈ N serves its

neighbors with the same exchange ratio (or, not at all).

• Any possible exchange equilibrium is also a lex-optimal allocation.

The competitive interactions of users embedded in a graph yield the same allocation point a central designer would have selected.

Stability wrt Coalitions

• Assume that subsets of nodes may decide to exchange only among

themselves.

• NTU Coalitional Service Exchange game:
• Played over the graph G = (N, E), by N players.
• Each node i has strategy di =
• dij : j ∈ Ni, P

j dij = Di

• , and utility ri.
• (Strong) Stability Definition:
• An allocation d (and the resource vector r) is called strongly stable if

∀S ⊆ N, there is no allocation b dS on the induced subgraph GS = (S, ES), such that b ri ≥ ri ∀i ∈ S, and b rj > rj for at least one node j ∈ S.

• Theorem: The only (strongly) stable allocations with respect to

coalitions are those with lexicographically optimal resource vectors r⇤.

• Hence, the solution of the graph-constrained coalitional game has the above

topological and price properties.

Dynamic Interactions

• How can the nodes find this equilibrium?
• Dynamic setup:
• Each node i creates ”service token” (e.g., relay opportunity) according to a

Poisson process with rate λi = Di.

• Every token is allocated to the neighbor with the lowest exchange rate (i.e.,

larger reciprocation).

• Decentralized and asynchronous best response under limited information.
• Extensive numerical results show that the system converges to the

unique vector of exchange ratios ρ⇤.

• Previous works showed convergence numerically for similar models, or

even proved it under certain conditions.

F . Wu, et al, Proportional Response Dynamics Leads to Market Equilibrium, ACM STOC’ 07

• B. Birnbaum, et al., Distributed Algorithms via Gradient Descent for Fisher Markets, EC’ 11
• R. Cole, et al., Fast-Converging Tatonnement Algorithms for One-Time and Ongoing Market

Problems, ACM STOC’ 08.

Overview

• The above models are motivated by the sharing economy.
• MNOs and start ups are already employing similar ideas.
• Novel opportunities for fundamental and experimentation-driven research.
• What will be the Uber, or Airbnb model for wireless networks?
• Mobile data offloading (leased) architectures.
• Leverage dormant user-owned capacity.
• Designed an efficient market mechanism.
• Move towards carrier-grade offloading solutions.
• UPN collaborative systems.
• Bottom-up networking solutions.
• Designed a resource allocation policy.
• Implemented a prototype system.
• Network resource exchange economies.
• Decentralized and dynamic bartering markets.
• Characterized the structure and efficiency of equilibriums.