Upload
isaac-cooper
View
217
Download
0
Tags:
Embed Size (px)
Citation preview
Modeling the Economics of Network Technology Adoption & Infrastructure Deployment
Soumya Sen
24th September, 2010. Princeton University
Research Motivation• Networked Systems have a ubiquitous presence
– e.g., Internet, Power grid, Facilities Management networks, Distributed databases
• Success of new network technologies depends on:– Technical advantage– Economic factors (e.g. price, costs, demand)
• Many technologies have failed– e.g., IPv6 migration, QoS solutions
• How to assess (design) new network technologies (architectures) for technical and economic viability?– Need for analytical frameworks– Need for a multi-disciplinary approach
224th September, 2010. Princeton University
Assessing Network Technologies
• Topic 1: – Network Technology Adoption/ Migration
• How can a provider help its technology (service) to succeed?
• Topic 2: – Network Infrastructure Choice
• What kind of network architecture should the new technology (service) be deployed on?
• Understanding Trade-offs between Shared and Dedicated networks
• Topic 3:– Network Functionality Richness
• How much functionality should the new network architecture have?
324th September, 2010. Princeton University
Research Contributions (1)
• Network Technology Adoption• Dependencies across users from network based interactions
(externality)• Incumbent’s have advantage of installed base• Technology gateways impact network externality, and hence
adoption
– Explored the dynamics of adoption as a function of user decisions
– Characterized the convergence trajectories and equilibrium outcomes
– Analyzed the role of gateways in technology migration
424th September, 2010. Princeton University
Research Contributions (2)
• Shared vs. Dedicated Networks• Many services on a common (shared) network vs.• Many services over separate (dedicated) networks
– Network choice depends on benefits of compatibility among offered services and demand uncertainty of new services
– Identified trade-offs and guidelines for network design
524th September, 2010. Princeton University
Network Technology Adoption
Topic 1:
624th September, 2010. Princeton University
Talk Outline:
1. Problem Formulation2. Model & Solution Methodology3. Key Findings & Examples4. Conclusions
Prior Work• Models that do not consider individual user utility:
• Fourt & Woodlock (1960) – constant hazard rate model• Bass (1969) - extension to include “word-of-mouth” effect• Norton & Bass (1987) - successive generation of technology
adoption
• Models that consider user utility function:• Cabral (1990) – only single technology adoption• Farrell & Saloner (1992) - homogeneous users• Choi (1996) - extended F&S to include converters• Joseph et al. (2007) – homogeneous users, doesn’t model
system dynamics
724th September, 2010. Princeton University
Problem Formulation
• Two competing and incompatible network technologies (e.g., IPv4 IPv6)– Different qualities and price– Different installed base
• Users individually (dis)adopt whichever technology gives them the highest positive utility
– Depends on technology’s intrinsic value and price– Depends on number of other users reachable (externality)
• Gateways offer a migration path– Overcome chicken-and-egg problem of first users
• Independently developed by each technology
– Effectiveness depends on gateways (converters) characteristics/ performance• Duplex vs. Simplex (independent in each direction or coupled)• Asymmetric vs. Symmetric (performance/ functionality wise)• Constrained vs. Unconstrained (performance/functionality wise)
824th September, 2010. Princeton University
A Basic User Model
• Users evaluate relative benefits of each technology
– Intrinsic value of the technology• Tech. 2 better than Tech.1 • denotes user valuation (captures heterogeneity)
– Externalities: linear in no. of users - Metcalfe’s Law• Possibly different across technologies (captured through β)• captures gateway’s performance
– Cost (recurrent) for each technology
924th September, 2010. Princeton University
IPv4 (Tech.1) IPv6 (Tech. 2)
10
Technology 1: U1(,x1,x2 ) = q1+(x1+α1β x2) – p1
Technology 2: U2(,x1,x2) = q2+(βx2+α2x1) – p2
– Cost (recurrent) of each technology (pi)
– Linear Externalities (Metcalfe’s law)• Normalized to 1 for Tech. 1• Scaled by β for Tech. 2 (possibly different from Tech. 1)
• αi, 0αi 1, i = 1,2, captures gateways’ performance
– Intrinsic technology quality (qi)
• Tech. 2 better than tech. 1 (q2 >q1)
– User sensitivity to technology quality ( ) • Private information for each user, but known distribution
24th September, 2010. Princeton University
Low-def. video (Tech.1) High-def video (Tech. 2)
• Low-def & High def video-conferencing service – Low-def has a lower price , but lower quality– Video is an asymmetric technology
• Encoding is hard, decoding is easy– Low-def subscribers could display high-def signals but not
generate them• Externality benefits of High-def are higher than those of Low-
def
• Converter characteristics– High/Low-def user can decode Low/High-def video signal– Simplex, asymmetric, unconstrained
1124th September, 2010. Princeton University
User Adoption Process
• Decision threshold associated with indifference points for each technology choice: 1
0(x), 20(x), 2
1(x) ,where x=(x1, x2)
– U1(, x) > 0 if ≥ 10(x) - Tech. 1 becomes attractive
– U2(, x) > 0 if ≥ 20(x) - Tech. 2 becomes attractive
– U2(, x) > U1(, x) if ≥ 21(x) - Tech. 2 over Tech. 1
• Users rationally choose– None if U1< 0, U2< 0– Technology 1 if U1> 0, U1> U2
– Technology 2 if U2> 0, U1< U2
• Decisions change as x evolves over time
12
x1 x2
24th September, 2010. Princeton University
Diffusion Model
13
• Assume a given level of technology penetration x(t)=(x1(t),x2(t)) at time t– Hi(x(t)) is the number of users for whom it is rational to adopt technology
i at time t (users can change their mind)
– At equilibrium, Hi(x*) = xi*, i {1,2}
– Determine Hi(x(t)) from user utility function
• Adoption dynamics:– Users differ in learning and reacting to adoption information
– Diffusion process with constant rate γ< 1
2,1 ,)(
itxtxHdt
tdxii
i
24th September, 2010. Princeton University
H1( x(t)) H2( x(t))
Solution Methodology• Delineate each region in
the (x1,x2) plane, where Hi(x) has a different expression
– There are 9 such regions, i.e., R1,…, R9
– Regions can intersect the feasibility region S 0 x1+x21 in a variety of ways
• This is in part what makes the analysis complex
– trajectories cross boundaries
P
Q
R1
R2
R3
R4
R5
R6
R7
R8R9
02
01
001
101
102
002
112
x1=1
x2=1
0
012
1424th September, 2010. Princeton University
Computing Equilibria & Trajectories
1524th September, 2010. Princeton University
Trajectories:
Solve Hi(x*) = xi*, i {1,2} in each region
Key Questions
• What are possible adoption outcomes?– Combinations of equilibria– Stable/ Unstable
• Adoption trajectories?– Monotonic vs. chaotic (cyclic)
• What is the role of gateways?– Do they help and how much?
1624th September, 2010. Princeton University
Results (1): A Typical Outcome
• Theorem 1: There can be multiple stable equilibria (at most two)
• Coexistence of technologies is possible – even in absence of gateways
• Final outcome is hard to predict simply from observing the initial adoption trends
1724th September, 2010. Princeton University
Results (2): Gateways may help Incumbents
• Theorem 2: Gateways can help a technology alter market equilibrium from a scenario where it has been eliminated to one where it coexists with the other technology, or even succeeds in nearly eliminating it.
• Gateways need not be useful to entrant always!• No gateways: Tech. 2 wipes out Tech.1 • Perfect gateways: Tech. 1 nearly wipes out Tech. 2
1824th September, 2010. Princeton University
Results (3): More Harmful Gateway Behaviors• Theorem 3: Incumbent can hurt its market penetration by introducing a
gateway and/or improving its efficiency if entrant offers higher externality benefits (β>1) and users of incumbent are able to access these benefits (α1β>1)
• Theorem 4: Both technologies can hurt overall market penetration through better gateways. Entrant can have such an effect only when (α1β<1). Conversely, Incumbent demonstrates this behavior only when (α1β>1)
Takeaway: Gateways can be harmful at times. They can lower market share for an individual technology or even both.
1924th September, 2010. Princeton University
Results (4): More Harmful Gateway Behaviors
• Theorem 5: Gateways can create “boom-and-bust” cycles in adoption process. This arises only when entrant exhibits higher externality benefits (β>1) than incumbent and the users of the incumbent are unconstrained in their ability to access these benefits (α1β>1)
Corollary: This cannot happen without gateways, i.e., in the absence of gateways, technology adoption always converges
Takeaway: Gateways can create perpetual cycles of adoption/ disadoption
P.S: Behavioral Results were tested for robustness across wide range of modeling changes
2024th September, 2010. Princeton University
Technology 1 Technology 2
Full-circle!
Limit Cycles: An Intuitive Explanationα1β>1
Technology 1: U1(,x1,x2 ) = q1+(x1+α1β x2) – p1 Technology 2: U2(,x1,x2) = q2+(βx2+α2x1) – p2
2124th September, 2010. Princeton University
Conclusions• Gateways can be useful to:
– Promote coexistence & improve market penetration– Help lessen price sensitivity
• But, Gateways can be harmful too:– Hurt an individual technology– Lower Overall Market– Introduce Market Instabilities
• Analytical model is useful in: – Identifying scenarios for policy intervention– developing long-term strategic vision
• Qualitative results are robust to: – switching costs– variation in utility function– non-uniform distr. of user preferences
2224th September, 2010. Princeton University
Network Infrastructure Choice:Shared Versus Dedicated Networks
Topic 2:
2324th September, 2010. Princeton University
Talk Outline:
1. Problem Formulation2. Model & Solution Methodology3. Key Findings & Examples4. Conclusions
Motivation• Emergence of new services require:
– Network provider has to decide between:• Common (shared) Network Infrastructure• Separate (dedicated) Network Infrastructure
• Examples:– Facilities Management services & IT
• e.g. IT & HVAC systems
– Video and Data services• e.g. Internet & IPTV services
– Broadband over Power lines
• Lack of Framework to evaluate choices:– Ad-hoc decisions (AT&T U-Verse versus Verizon FiOS)– Manufacturing Systems Literature:
• Plant-product allocation, optimal resource allocation
2424th September, 2010. Princeton University
Problem Formulation• Two network services (technologies)
– One existing (mature) service – One new service with demand uncertainty
• Costs show economies or diseconomies of scope
• New service has demand uncertainty– Needs capacity provisioning
• before demand gets realized
– Dynamic resource “reprovisioning”• But some penalty will be incurred (portion of excess demand is lost)
– Technology advances allow Reprovisioning (e.g., using virtualization)
• How critical is reprovisioning ability in choosing network design?– Compare networks based on profits
2624th September, 2010. Princeton University
Model Formulation
• Basic Model: A Two-Service Model
• Service 1 (existing service)
• Service 2 (new service with uncertain demand)
• Three-stage sequential decision process
• Compare Infrastructure choices based on expected profits
27
Reprovisioning Stage
Capacity Allocation Stage
Infrastructure Choice Stage
Solve backwards
24th September, 2010. Princeton University
Model Variables• Provider’s profit depends on:
– Costs:• Fixed costs• Variable costs
– grows with the number of subscribers (e.g. access equipment, billing)• Capacity costs
– incurred irrespective of how many users join (e.g. provisioning, operational)
2824th September, 2010. Princeton University
Cost Component Service 1 separate
Service 2 separate
Common
Fixed Costs cs1 cs2 cc
Contribution Margin
(grows with each unit of realized demand)
ps1 ps2 pc1, pc2
Variable Costs
(incurred irrespective of realized demand)
as1 as2 ac1, ac2
Gross Profit Margin= pi-ai (i={s2, d2})
Return on capacity= pi /ai
Solution (1): Reprovisioning Stage• Service 2 revenue: (i={s2, d2} for Shared and Dedicated respectively)
– when D2<Ki:
– when D2>Ki:– Reprovisioning Ability:
• A fraction “α” of the excess demand can be accommodated
User Contribution
Capacity cost
2924th September, 2010. Princeton University
• A word about reprovisioning ability,– Independent of the magnitude of excess demand– Captures feasibility of and latency in securing additional resources
– So what do and mean?
Solution (2): Capacity Allocation Stage
• Expected Revenue, E(Ri|Ki), for a given provisioned level Ki:
• Optimal Provisioning Capacity (for demand distribution ~U[0, D2
max]):
3024th September, 2010. Princeton University
Solution (3): Infrastructure Choice Stage• Dedicated Networks:
– Service 1 revenue:– Service 2 revenue under optimal provisioning:
– Total profit:
• Shared Network:
• Infrastructure Choice: – Common if , else separate
31
Profit from Service 2Profit from Service 1
24th September, 2010. Princeton University
Choice of Infrastructure
• Impact of system parameters:– Varying cost parameters affect the choice of infrastructure
• Shared to Dedicated (or Dedicated to Shared)• Single threshold for switching n/w choice
– Surprisingly, ad-hoc “reprovisioning” ability also impacts in even more interesting ways!
• Common is preferred over separate when
Independent of provisioning decision
Depends on provisioning decision
32
Diff. in optimal capacity cost
24th September, 2010. Princeton University
h(α)=
Function of pi, ai, α, i={s2,d2}
Analyzing the effect of α on h(α)• Proposition 1: Increase in α benefits both shared and dedicated networks.
(i) if ( ), increases in α benefits shared (dedicated) n/w more than dedicated (shared)
(ii) if ,( ), increases in α benefits shared (dedicated) more at low α and dedicated (shared) more at high α
• The value of h'(0) and h'(1) fully characterize the shape of h'(α)
3324th September, 2010. Princeton University
Gross Profit MarginReturn on Capacity
Results: Impact of Reprovisioning
3424th September, 2010. Princeton University
Some Design Guidelines
• Identify cost components • use the model to investigate the net economies/
diseconomies they create– Single threshold for switching choices for most cost parameters
• Check the impact of reprovisioning– Whether α has an effect depends on
• The sign of the derivative h'(α)• Use the two metrics to identify operational region• The magnitude of γ (how far from zero)• Outcomes: Zero, one or two transitions
3524th September, 2010. Princeton University
Conclusions
• Developed a generic model captures economies and diseconomies of scope between shared and dedicated networks
• Reprovisioning can affect the outcome in non-intuitive ways– Validates the need for models to incorporate this feature– Yields guidelines on how reprovisioning affects choice of
architecture
• Identified key operational metrics to consider– Provided decision guideline
3624th September, 2010. Princeton University
Ongoing Work & Future Extensions
• Strategic selection of gateways in network technology adoption
• Dynamics of adoption in two sided markets
• Understanding trade-offs between minimalist and functionality-rich network architectures
3724th September, 2010. Princeton University
Bibliography(1) S. Sen, Y. Jin, R. Guerin and K. Hosanagar. Modeling the Dynamics of Network Technology
Adoption and the Role of Converters. IEEE/ACM Transactions on Networking. 2010
(2) S. Sen, Y. Jin, R. Guerin and K. Hosanagar. Technical Report: Modeling the Dynamics of Network Technology Adoption and the Role of Converters. Technical Report. June, 2009. Available at http://repository.upenn.edu/ese papers/496/.
(3) Y. Jin, S. Sen, R. Guerin, K. Hosanagar and Zhi-LiZhang. Dynamics of competition between incumbent and emerging network technologies. In Proc. Of ACM NetEcon'08, pp.49-54, Seattle, 2008.
(4) S. Sen, R. Guerin and K. Hosanagar. Shared Versus Separate Networks - The Impact of Reprovisioning. In Proc. ACM ReArch'09, Rome, December 2009.
(5) S. Sen, K. Yamauchi, R. Guerin and K. Hosanagar. The Impact of Reprovisioning on the Choice of Shared versus Dedicated Networks. Submitted to WEB, December 2010.
(6) R. Guerin, K. Hosanagar, S. Sen and K. Yamauchi. Shared versus Dedicated Networks: The Impact of Reprovisioning on Network Choice. Under preparation for INFORMS journal on Information Systems Research.
Acknowledgements: Roch Guerin (ESE, Penn), Kartik Hosanagar (Wharton, Penn), Y. Jin (ESE, Penn), Kristin Yamauchi (ESE, Penn), Andrew Odlyzko (Math, UMinn), Zhi-Li Zhang (ECE, UMinn)
3824th September, 2010. Princeton University
Thank You!