Interconnection:
An Economic Perspective
Peyman Faratin (CSAIL)
Steven Bauer (CSAIL)
David Clark (CSAIL)
Bill Lehr (CSAIL)
Arthur W Berger (Akamai,CSAIL)
Patrick Gilmore (Akamai)
Tom Wilkening (Economics)
Interconnection Problem
•
•
•
•
•
•
•
•
•
AT&T - Carter phone & Hush-a-Phone (blocking)
….
2002: Madison-River - Vonage (blocking)
2005: Cogent-Level 3 (disconnecting)
2006: AT&T - Google (tiering)
2007: T-Mobile2 (blocking)
…
ICE (Farrell and Weiser), Agency (Milgrom et.al),
Entry Story -- because of lack of quality competition in interconnection
 Two-Sided Markets (New Institutionalist Model)
 A model of value-flows - demand information
 Market failures
• “middlebox”/overlays entry
• Interconnection discrimination incentives (given cost-allocation mechanism)
Industrial Organization:
Two-Sided Markets
• Generative: Design aid
• Business Model
• Descriptive: future regulatory thinking
Causal Hypothesis of
Interconnection Problems
Outcomes
Information & Behaviors
IO & Contracts
Architecture
The Trinity:
Institution, Strategies and Outcomes
Institution
• architecture
• contract
• policy
Strategic
Agents
Outcomes
• Scalability, Resilience, Convergence
• Fairness, Innovation, Profitability
Transfer Distribution Ambiguities
(“we know how to route packets but not money”)
AS1
AS1
$
$
AS2
AS2
content
Ambiguities Galore
content
AS1
AS1
AS1
$
$
$
AS2
AS2
AS2
content
Solution:
Bi-lateral Volume-Based Contracts
IO & Contracts
• Retail market (bursty):
 Flat-rate
 Peak-rate tiered pricing
Architecture
• Wholesale market (better aggregation “deeper in”):
 Full transit
• Transfer level = non-linear
• Transfer structure = asymmetric
 Peering
• Transfer level = 0
• Transfer structure = N/A
 Emerging mechanisms: Paid-peering & Partial Transit
• Distribution of Fixed and Usage pricing
End-Hosts Bear Cost of
Transport
AS
$=0
$
AS
$
MIT
AS
$
$=0
AS
$
http://www.google.com
No E2E Accounting for Tastes
Eyeballs
$
ISP
$$$$
Growing
WebServer
Eyeballs
$$$$
ISP
$$
Established
WebServer
Eyeballs
?$
ISP
?$
Public
WebServer
Coordination Failures Has Led
to E2E Market-Failures
CU i
Pj1
Pi1
CP j
IAP1
P12
CU k
Pk2
IBP2
Pm2
CP m
Market-Failure Induced CDN Entry
AKAM: 20,000 servers,900 networks,70 countries,750 cities, serving ≈ 15% of content
CU i
Pj1
Pi1
IAP1
P1
P12
PjCDN
PC2
CDN
IBP2
Pj2
P22
CU k
CP j
PmCDN
P2
Pm2
Pk2
IAP2
CP m
Strategies and Outcomes
Outcomes
Information & Behaviors
Contracts
The Trinity:
Institution, Strategies and Outcomes
Strategic
Agents
Institution
• architecture
• contract
• policy
Outcomes
• Fairness
• Growth
• Profitability
Who Should Pay Who?
Primitive = Value-Flows
i
Pi
eyeball
Pj
ISP
pj
“Free Goods” IV
I
(0,0)
III
II
j
Content
Provider
Q: what is the optimal
price structure for ISP
to maximize profits?
pi
Value-Flow Discrimination
i
eyeball
Pi
ISP
Pj
j
Content
Provider
Q: what is the optimal price structure?
A: Depends on:
• Relative size of value flows (cross-market externalities)
pj
• Fixed / Per transaction prices
• Single v.s Multi-homing
45o
Established commercial
web-server  $$  ISP
 $$$  eyeballs
pi
Complementarities/Interactions:
Multi-Product Markets
Value-Flows/Externalities:
Chicken-Egg Problems
Two-Sided Markets
• But platform has to solve “chicken-egg” Problem: if there were
more women, then more men would come, more women would
come, more men would come,….  discrimination is welfare
enhancing. “ladies nights”
Non-Discrimination Institution
Strategic
Agents
Institution
• “no ladies night”
Outcomes
• Fairness
• Growth
Does Institution Implement
Desired Outcome?
• Rule (motivated by “fairness”): No bars can
access discriminate based on sex
• Q: Does rule implement a “fair” & innovative
outcome in the presence of strategic actors?
• A: No. Institution is “fair” but gives no growth
incentives. Neutrality rule is not neutral with
respect to growth  tussle between objectives
Result of Rule: Closes Some
Markets, Others Grow but
Inefficiently
Strategic Preferences of Content
Providers & Users
Eyeballs
$
ISP
$$$$
Growing
WebServer
Eyeballs
$$$$
ISP
$$
Established
WebServer
Eyeballs
?$
ISP
?$
Public
WebServer
Strategic Agent Preferences:
The Platform (in Presence of
Externalities)
• Platform (ISP/CDN) solves for efficient prices:
 market price level ( P *) and
*
*
 price structure [Pi ,P j ]
• Profit maximizing pricing structure in
presence of externalities is often

discriminatory (subsidize one side of the

market to stimulate demand on other side c.f. bar)
 Strong incentives to discriminate
Network Neutrality Law or Current
Architecture & Protocols
Institution
• “the architecture
can’t / shouldn’t
do that”
• “no price discrimination
for same service”
Strategic
Agents
(1:Customer, 2:Content Provider)
Outcomes
• Fairness
• Growth
3: Platform:
ISP
Unintended Outcome of
Institution: Market Closures
Eyeballs
$
ISP
$$$$
Growing
WebServer
Eyeballs
$$$$
ISP
$$
Established
WebServer
Eyeballs
?$
ISP
?$
Public
WebServer
Externalities Create Surplus
Expansion Opportunities (v.s.
Capture)
• Traditional (one-sided) Price discrimination
 Discrimination increases the profits of the
monopolist but may open some markets
that would otherwise be closed.
• … platform intermediaries in a TSM seek to
maximize profit by transferring surplus from
seller to consumer thereby growing the market
 Growth on one side of the market induces
growth on the other, creating surplus that
can be captured
Market-Failure Induced CDN Entry:
Akamai: 20,000 servers, 900 networks, 70 countries, 750 cities, serving ≈ 15% of
content
CU i
Pj1
Pi1
IAP1
P1
P12
PjCDN
PC2
CDN
IBP2
Pj2
P22
CU k
CP j
PmCDN
P2
Pm2
Pk2
IAP2
CP m
Architectural Tools We Provide
• The real question is how to architect for it:
 Change in demand in i market /
change in demand in j market
 Source-destination discrimination
 App discrimination
 Per packet/per flow bit discriminate
 Encryption
 ….
• There is a delicate tradeoff involved in
how much information we provide and
how much we lose/gain in objectives we
are interested in
Outcomes
Information & Behaviors
IO & Contracts
Architecture
Conclusion
• Interconnection




Not only a L2, L3 problem
Contract engineering and value-flows
Agents use mechanisms strategically
Tussle over outcomes
• Open Questions:
 Preferences over outcomes/objectives
 CDN Tipping and Market-Power
• 2 tiered Internet?
 Externality Information for monitoring and regulation
• Industrial Organization
 A tool for architecture & policy
Future: ICWG
• Data
 War Stories/cases
•
•
•
•
Peering of video
Exclusivity contracts
Games being played
….
 Quantities and prices
  data to support theory
  data to build theory
• Informative process to all
 Designers
 ISPs
 Policy makers
Peyman@mit.edu
Auxiliary Slides (I)
Information and Strategic Games
Competition: Peering+Transit
Strategic Interactions
• All compete to:
 establish and
 maintain peering
• Competition over:
 Eyeball Networks
 Content
• Colo CP (Apple iTunes, Microsoft,..)
• Stub ASs (Yahoo, Google,…)
• Non-stub Tier2 content (transit providers to content Stub AS)
“Normal” Business Strategy of LE-LC
Strongest Peering Incentives
LE
• Assume LE-LC interconnect under peering
• LC’s problem is to keep ratios
LC
LE-LC Strategies
LE
LC
• Observations:
 Eyeballs are fixed, content can move (switching costs of content is
lower)  perception of bargaining power by LE
 LE doesn’t care about being out of balance & in fact wants to be out
of ratios so it can demand payments (paid-peering)
“Equilibrium” in Establishing New
Peering between Strategic
Networks
A < E, B > F
G > C, H < D
P
LE
P
LC
P
P
(A,B)
(C,D)
(E,F)
(G,H)
LE-LC Peering Establishing Strategies
LE
LC
• LE strategy:
 LC asks to peer (or upgrade peering facilities to keep abreast
of traffic flows)
 LE refuses and demands higher settlements (paid-peering)
because:
• it is LC who is out of ratios and causing costs
• Operational costs (AOL)
• Precedence settings leads to economic loss on the long-run
 Most LCs refuse to pay, but some do concede. Some content
owners on LC who doesn’t concede switch to LCs that do.
LE-LC Strategies: Vertical + Horizontal
tier1
tier1
P2
P3
P1
LE
•
LC’s Counter strategy (“chicken”):


•
LC
If LE refuses to peer/upgrade peering then LC sends some traffic via transit
Punishing strategy: LC bears P2 (which may even be above cost of P1), but LE has to pay P3
Condition: Strategy only works if both LC&LE are transit customers of tier1. If LE has
peering with tier1 & LC sent via transit then LC would in fact be helping LE because LE
would look bigger to tier 1
LC’s Strategy to Keep Ratios: Sell
Low-cost Transit (Poaching:
Vertical+Horizontal)
LE
LC
P2
SE
•
LC’s strategy:




Peering link is full-duplex and LC is mostly outbound
To keep ratios LC needs to pull  sell transit to SE
Poaching SEs by setting P2 at or even below cost
 LE P2P traffic to SE goes via LC
LC’s Strategy is Reactive and
Proactive
LE
T
SE
LC
P2
SE
SE
Ratio Balancing Needs Create Poaching
Competition, Downward Pressure on
Transit Prices and Quality
LC
LC
LC
LE
P2
SE
• Margins of gain of poaching strategy to maintain peering shrinks
as P2 falls
• Excess reductions of P2 lowers quality/performance of transit
because incentives of LC to manage are eroding?
Salient Economic Features
• Dynamic efficiency (innovation)
• Operator IO is highly complex (no clear
upstream/downstream)
• Behavioral:
 Direct & indirect network Effects
 Unobservability
 Coordination failures
Auxiliary Slides (II)
TSM Model
How ISP Determines its
Optimal Price Structure:
Geometry of the Problem
i
Pi
eyeball
Pj
ISP
pj
“Free Goods” IV
I
(0,0)
III
j
Content
Provider
Q: what is the optimal
price structure for ISP
to maximize profits?
pi
II
Value-Flow Discrimination
i
eyeball
Pi
ISP
Pj
j
Content
Provider
Q: what is the optimal price structure?
A: Depends on:
• Relative size of value flows (cross-market externalities)
pj
• Fixed / Per transaction prices
• Single v.s Multi-homing
45o
Established commercial
web-server  $$  ISP
 $$$  eyeballs
pi
Total Consumption
qi  Di ( pi )  e jiD j ( p j )
i’s “native” demand

demand of i due to
demands of j
Total Consumption
qi  Di ( pi )  e jiD j ( p j )
q j  D j ( p j )  eij Di ( pi )
q i
e ji 
q j
network externality term (how much
purchases in j market affects
purchases in the i market)
Benchmark: eji = eij = 0
pj
pi(pj)
Po = (1/2,1/2)
pj(pi)
1/2
1/2
pi
eij 
pj
q j
q i
 1/ 3
pj
eji=3/4
eji=0
pi

pj
eji=11/10
pi(pj)
pj(pi)
pi
Architectural Guide
• eij a potential candidate for value-flow
proxy
Value-Flow and Structural
i
usr
Pi
ISP1
j
Pj
Google
usr
ISP2
pj
Q: what is the optimal price structure?
A: Depends on:
• Relative size of cross-group externalities
• Fixed / Per transaction prices
45o
• Single v.s Multi-homing
pi
Assumptions
• Network’s tariff:
 Charges to i market for subscription
 Charges to j market for traffic termination
• i market single-homed
 Makes single either-or decision  competition between
platforms for i market
 i chooses network that maximizes its surplus
• j market multi-homed
 Makes independent join decisions  no competition
between platforms for j market
 j puts more weight on network benefits of being in contact
with widest population of i market than transaction costs of
multiple platforms
Equilibrium Tariff (M. Armstrong)
pj
IV
I
pi
III
II
• Low subscription charges to i market and high
termination charges to j market
 Equilibrium termination charges to j market maximizes i
market and network’s profits and ignores j market welfare.
Multi-homing Reduces
Competition and Welfare
• Single-homing side is treated well, mhoming side’s interest are ignored at
equilibrium (i is even cross-subsidized)
• “Competitive bottleneck”: even if market
for content users is highly competitive,
so that profits of networks are lowered,
there is no competition for providing
services to content providers.
Engineers Provide Tools to
Firms: Design-Evaluate Cycle
• IO methodology:
 puts economics (back) into the design
consideration, but after protocol design
 Allows “comparative statics” - “what happens to
welfare if we change the institution”
 Build testable models to ask “what-if” questions on
efficiency-fairness tradeoff
Future
• Competition for ideas and incentives
 Strategic agents will use technical & regulatory
tools to their economic advantage
• FIND (2006):
 3/10 economic (CABO, Virtualization, Architecture
of all fiber networks)
 Highly recommend talking to economists &
regulators
 SIGCOMM 08 Workshop?
• MIT’s Interconnection Working Group
 David Clark, Steven Bauer, Bill Lehr, Peyman
Faratin, Akamai
Markets
i
Pi
usr
Pj
Google
ISP
qi
j
qj
Geometry of the Price
Discrimination Problem
i has relatively more externality
impact on j
p
j
MC
pi
j has relatively more externality
impact on i
Demand
• Each market has a continum of
consumers willing to buy one discrete
unit of good (transport service)
• Let v be arbitrary willingness to pay of
an individual
• Then D(p) is the market demand
CU’s Market Demand
Maximum
service
value (in absence
of network externalities)
D(pi)
Vi
Maximum
market
size (in absence
of network externalities)
v
vp
dDdpi
Qi
Total Consumption
qi  Di ( pi )  e jiD j ( p j )
i’s “native” demand

demand of i due to
demands of j
Total Consumption
qi  Di ( pi )  e jiD j ( p j )
q j  D j ( p j )  eij Di ( pi )
q i
e ji 
q j
network externality term (how much
purchases in j market affects
purchases in the i market)
Measures
q j  D j ( p j )  eij Di ( pi )
q j
'
 eij Di ( pi )
pi
Spill-over/TS network
externality =
cross-price (i to j)
contribution to sales
in j market.

Measures
qi  Di ( pi )  e jiD j ( p j )
q j  D j ( p j )  eij Di ( pi )
q j

pi
r
 qi
p j
Importance of the
spill-over effects
Externality of CPs to CUs
• As CPs use more transport then CUs
max. service value for transport
increases
D(pi)
qi  Di ( pi )  e jiD j ( p j )
Vi

Qi
Externality of CPs to CUs
• CU value increase  Di1(e jiD j )
Vi

Qi
qi  Di ( pi )  e jiD j ( p j )

Consumer Surplus
qi  Di ( pi )  e jiD j ( p j )
Vi
Vi
Sji = (eji Qj)Vi /2
= QiVi / 2
S
i = QiVi / 2
Qi
Qi
Surplus Division v.s. Capture
• Third-Degree Price discrimination
 Firms offer nonlinear prices to mixed markets
force heterogeneous consumers to self select
(Peak-rate pricing?)
 Mechanism differentially extract consumer surplus
and transfer it to the seller
• … platform intermediaries in a TSM seek to
profit by transferring surplus from seller to
consumer
 Growth on one side of the market induces growth
on the other, creating surplus that can be captured
Monopoly Pricing in Absence of
Network Externality (Po)
(monopoly sets prices in the two
markets independently, implicitly
assuming eij = eji = 0)
Monopoly CUs Profits
 i  piQi (1
Vi
Vi / 2
Vi o Qi
p  ,qi 
2
2
 io
Qi / 2
Vi
o
i
(qoi , poi )

pi
Qi
ViQi Si
 

4
4
o
i
)
Monopoly Pricing with Network
Externality (P*)
(monopoly sets prices in the two
markets interdependently, eij  eji > 0)
• Assume:
 j market (CPs) demand for transport is
inelastic
 i market (CUs) demand for transport is
elastic
 eji  eij > 0
 qi / pj > 0 (Positive TS, spillover, effect)
qi  Di ( pi )  e jiD j ( p j )
q j  D j ( p j )  eij Di ( pi )
-
eji  eij > 0, i > j
pi
+
-
pj
qi=Di(pi)
+
qj=Dj(pj)
+
+ q =e D (p )
j
ij i i
qi=ejiDj(pj)
+
-
pi
pi
+
qj=Dj(pj)
pj
qi
+
qi=Di(pi)
pj
+
+ q =e D
j
ij i(pi)
qi=ejiDj(pj)
pj
+
qj
pi
qi
pi
qi
Asymmetricity in Externalities
• Now vary the relative influence of CP
CU
–eji  eij > 0
Benchmark: eji = eij = 0
pj
pi(pj)
Po = (1/2,1/2)
pj(pi)
1/2
1/2
pi
pj
pi(pj)
II
I
pj(pi)
1/2
pi
1/2
III
IV
eij = 1/3
pj
pj
eji=3/4
eji=0
pi
pj
eji=11/10
pi(pj)
pj(pi)
pi