managing qos
DESCRIPTION
QoSTRANSCRIPT
-
Managing QoS in a Multiprovider Network
Jean Walrandhttp://www.eecs.berkeley.edu/~wlr
University of CaliforniaBerkeley
Thanks to Jrn Altmann, Linhai He, and Jun Shu for their help.
-
2Jean Walrand Porquerolles May 03
QoS in a Multiprovider Network Overview Approaches Social Welfare & Packet Marking Market Segmentation Games: Mini-Tutorial Pricing Summary Related Topics References
TOC
-
3Jean Walrand Porquerolles May 03
Overview Internet Structure Implications Different Views Interactions Peering Agreements
TOC OTOC verview
-
4Jean Walrand Porquerolles May 03
Internet Structure
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
Domains Owned and operated by
independent organizations
Have own set of policies
Relationship between domains Customer and provider Collaborators and
competitors
TOC Overview Structure
-
5Jean Walrand Porquerolles May 03
Internet Structure Backbone - AT&T
http://www.caida.org/tools/visualization/mapnet/Backbones/
TOC Overview Structure
-
6Jean Walrand Porquerolles May 03
Internet Structure Backbone - Sprint
http://www.caida.org/tools/visualization/mapnet/Backbones/
TOC Overview Structure
-
7Jean Walrand Porquerolles May 03
Internet Structure Autonomous Systems
http://www.caida.org/analysis/topology/as_core_network/pics/ascoreApr2002.gif
TOC Overview Structure
-
8Jean Walrand Porquerolles May 03
Internet Structure
[GH00]
Transit Customer Peering TOC Overview Structure
-
9Jean Walrand Porquerolles May 03
Network Structure
D EC F
BA
AC {1,2,3}AD {4,5}
D{4,5}DC{1,2,3}
BAC{1,2,3}BAD{4,5}
C{1,2,3}
C{1,2,3}
BGP
1 2 3 4 5
A announces preferred paths: Policies E.g., A prefers C{1,2,3} to DC{1,2,3} because shorter
TOC Overview Structure TOC Overview Structure
-
10Jean Walrand Porquerolles May 03
Implications
End to End QoS Requires collaboration among all domains Proposed solutions should provide right
economic incentives Scalability
Flow vs aggregate based traffic management
Techniques to eliminate states inside networks
TOC Overview Implications TOC Overview Implications
-
11Jean Walrand Porquerolles May 03
Different Views End Users
Have utility for services Demand, Willingness to pay
Network as a Public Good Maximize total utility: Social
Welfare Providers as Market Players
Maximize Revenues
TOC Overview Different Views TOC Overview Different Views
-
12Jean Walrand Porquerolles May 03
Interactions Agents
ISPs, Users Network Leads, Users Follow
Network Set Prices p; Users Respond D(p) Auction
Users Bid, Network Set Prices, Users Respond Many ISPs
Non-cooperative: Each ISP j sets price pj Cooperative: ISPs agree on prices, share revenues
What is the product? Single class or multiple classes of service
TOC Overview Interactions TOC Overview Interactions
-
13Jean Walrand Porquerolles May 03
Peering AgreementsCaricature
N4 N5N3 N6
N2N1
A B C D
N1, N2 are Tier-1 Networks N3 N6 are Tier-2 Networks A D are clients
TOC Overview Peering
-
14Jean Walrand Porquerolles May 03
Peering Agreements
N4 N5N3 N6
N2N1
A B C D
N3, N4 pay N1 for traffic: mean, peak, (N3 is a client of N1; N1 is the provider of N3)
N1 N2 may agree to peer
TOC Overview Peering
-
15Jean Walrand Porquerolles May 03
Peering Agreements
N4 N5N3 N6
N2N1 RPeering
Customer
N1 and N2 Peer:Free transport of each others trafficShare cost of R
TOC Overview Peering
-
16Jean Walrand Porquerolles May 03
Peering Agreements
R
N4 N5N3 N6
N2N1 RCustomer
Should N3 and N4 peer?+: Reduce their transit cost through N1-: If |N3| >> |N4|, then N4 free rides N3-: If N3 is faster then N4, then peering might entice
customers of N3 to join cheaper N4 instead. Large ISPs dont like to peer with small ISPs
[BW98][CRT00]
TOC Overview Peering
-
17Jean Walrand Porquerolles May 03
Peering Agreements - BGP
C
A
B DE
E{..}
E{..}
E{..}E{..}
E{..}
F
GE{..}
By not announcing the paths E{..} to D, A limits its transittraffic A favors its paying customers G, at the expenseof the customers F of D
TOC Overview Peering
-
18Jean Walrand Porquerolles May 03
Peering Agreements - BGP
E
E{..}
E{..}
E{..} E{..}E{..}
E{..}
F
G
Selfishly, ISPs may have a benefit in not announcing some routes. However, if they all do this
TOC Overview Peering
-
19Jean Walrand Porquerolles May 03
Peering Agreements Note: BGP is blind to utilization
E F
G
This inefficient routing is not a big problem if utilizationis very low; which is fine when there is vast over provisioning.
However, protection switching also adds to the over provisioning.All this adds up to a big waste .
TOC Overview Peering
-
20Jean Walrand Porquerolles May 03
Approaches What and Where to Control? Best Effort Per Flow Reservation Per Flow End-to-End Admission Edge-to-Edge Pipes Local Pipes Some Lessons
TOC Approaches
-
21Jean Walrand Porquerolles May 03
What and Where to Control? End-to-end flow, between transport ports Across the network, between edge routers Across each domain Across each individual link Granularity
TOC Approaches What and Where to Control? TOC Approaches What and Where to Control?
-
22Jean Walrand Porquerolles May 03
Best Effort
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
Great Scaling; Poor QoS
Packets competeDropped if congested
TOC Approaches Best Effort TOC Approaches Best Effort
-
23Jean Walrand Porquerolles May 03
Per Flow Reservation
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
Reserve B Mbps?
OK!
Poor Scaling and Response Time; Flow QoSTOC Approaches Per Flow Reservation
-
24Jean Walrand Porquerolles May 03
Per Flow Reservation - IntServBasic Observation [PG93] WFQ at each router + LB per flows
bounded delays Idea: Reserve (,, w) along the path : RSVP
TOC Approaches Per Flow Reservation
-
25Jean Walrand Porquerolles May 03
Per Flow Reservation - IntServ
Admission Control
C
o
n
t
r
o
l
P
l
a
n
e
D
a
t
a
P
l
a
n
e
Scheduler
Routing RSVP messages
Classifier
RSVP
Route Lookup
Forwarding Table Per Flow QoS Table
Routing Messages
Data InData Out
[Abhay Parekh]
TOC Approaches Per Flow Reservation
-
26Jean Walrand Porquerolles May 03
End-to-End Per Flow Admission
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
Test B Mbps?
OK!
Good Scaling; Unpredictable admission probability
Once admitted, guaranteed
TOC Approaches E2E Per Flow Admission
-
27Jean Walrand Porquerolles May 03
End-to-End Per Flow Admission A plug for virtual queues!Mark packets when utilization exceeds 90%, using virtual queues:
C
0.9C
With this scheme, we detect congestion before ithappens. The queues never really build up. [CK95, KS01]
TOC Approaches E2E Per Flow Admission
-
28Jean Walrand Porquerolles May 03
End-to-End Per Flow Admission Goal: Admitted calls have guaranteed QoS
(E.g., VoIP, video conference, .)
Scheme:Send test packets, say for 5 seconds. Routers mark test packets if utilization > 90%. (VQ)If no test packet gets marked, the connection is accepted.Otherwise, connection is blocked and goes away.Advantages:+ New calls do not disrupt accepted calls.+ No complex RSVP-like protocol; no state in routers.
TOC Approaches E2E Per Flow Admission
-
29Jean Walrand Porquerolles May 03
Edge-to-Edge Pipe (e.g., MPLS)
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
Complex [QoS routing]; Controllable QoS
Reserve big pipesPipes used by flows
TOC Approaches Edge-to-Edge Pipes TOC Approaches Edge-to-Edge Pipes
-
30Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ)
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
Simple [local]; Issue: How to provision SLAs?
Reserve local pipesPipes used by flows
TOC Approaches Local Pipes
-
31Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ)DiffServ
IngressEgressEgress
IngressEgressEgress
DS-1 DS-2
Kinds of Service:1. Premium2. Assured3. Best Effort
Kinds of Service:1. Premium2. Assured3. Best Effort
Edge router Core router[Abhay Parekh]
TOC Approaches Local Pipes
-
32Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ)DiffServ Example - VoIP 80 bytes every 20ms Give high priority to VoIP Assume VoIP utilization of a link is less than 50%
Then, delay of VoIP is less than 10ms per link(assuming preemptive priority add a few ms)
Indeed: Cycles of 20ms with no carry over Worst case: wait for back-to-back VoIP 10ms
TOC Approaches Local Pipes
-
33Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ) Provisioning SLAs: Worst Case is Bad!
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
1010
1030?
TOC Approaches Local Pipes
-
34Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ)Provisioning SLAs Can be slow Can have loops: AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
TOC Approaches Local Pipes
-
35Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ)Measurement-Based:
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
Measure LoadAdjust SLA preemptively
Good scaling and response time, but .TOC Approaches Local Pipes
-
36Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ) Measurement-Based:
AT&T
Sprint WorldCom
SBC
Level3
BBN
Stanford
UIUC
MIT
Ameritech
Berkeley
SLAs may have to be readjusted sub-optimally
S OK
S no longer OKS
TOC Approaches Local Pipes
-
37Jean Walrand Porquerolles May 03
Local Pipes (SLAs; e.g., DiffServ) Measurement-Based: Case StudyNew
Admit if peak(new) < Gap
Capacity
GapMean + 2.4
Multi-domain: Need a broker per domain, with negotiationsof admission quotas among domains .
TOC Approaches Local Pipes TOC Approaches Local Pipes
-
38Jean Walrand Porquerolles May 03
Some Lessons Comparison Why not much QoS?
TOC Approaches Lessons
-
39Jean Walrand Porquerolles May 03
Comparison
Per flow steupLong term setupNo setupComplexity
End-to-endDomain End-to-endService scope
Not scalable (each router maintains per flow state)
Scalable(edge routers maintains per aggregate state; core routers per class state)
Highly scalable (nodes maintain only routing state)
Scalability
Per flow isolationPer flow guarantee
Per aggregate isolationPer aggregate guarantee
ConnectivityNo isolationNo guarantees
Service
IntservDiffservBest-Effort
[Abhay Parekh]
TOC Approaches Lessons Comparison
-
40Jean Walrand Porquerolles May 03
Comparison It is a matter of scale
SF
UCB
Boston
BU
2Mbps
PoP
< 12 Mbps
Total < 8 MbpsUCD
MIT3Mbps
VoIP traffic UCB PoP to 2Mbps: MPLS pipe, etcTotal VoIP traffic at PoP: 8 Mbps: DiffServ, etc.Given access constraints, some regularity in bboneand some regularity in S/D pairs of VoIP .
TOC Approaches Lessons Comparison
-
41Jean Walrand Porquerolles May 03
Why not much QoS? Answer 1: Complexity, Scalability, .
Answer 2: Need agreements QoS must be end to end, i.e., across multiple providers. If ISP2 does not do it, why should ISP1 do it, and vice versa .
Answer 3: What are the economic incentives?Need pricingChargingRevenue Distribution
TOC Approaches Lessons No QoS
-
42Jean Walrand Porquerolles May 03
Social Welfare and Packet Marking We consider a congested link Key idea: Packet marking can provide a signal that decouples a social welfare optimization problem. We then see how a version of TCP can solve the decoupled problem Social Welfare Problem (Primal) Decoupled Problem (Dual) Example TCP ImplementationWhy does it work? Extension Virtual Queues
TOC Social TOC Social
-
43Jean Walrand Porquerolles May 03
Primal Consider the following problem:
TOC Social Primal
x Max U1(x) + U2(y)
Subject to x + y C
Max U1(x) + U2(y)
Subject to x + y CC
ySolution:The solution (x*, y*) is such that x* + y* = C.Also, U1(x* + ) + U2(y* - ) U1(x*) + U2(y*), for all .Hence,
U1(x*) + U2(y*) + {U1(x*) U2(y*)} U1(x*) + U2(y*).Since is arbitrary, this implies
U1(x*) = U2(y*) =: p
-
44Jean Walrand Porquerolles May 03
Dual Consider the following problem:
x Max U1(x) + U2(y)
Subject to x + y C
Max U1(x) + U2(y)
Subject to x + y CC
ySolution (continued):U1(x*) = U2(y*) =: p is equivalent to
User 1 maximizes U1(x) pxUser 2 maximizes U2(y) py
In this decoupled problem p is the price per unit bandwidth.Price p is a congestion price that reflects the impact of anadditional unit of bandwidth on the utility of the other user.[p is a.k.a. shadow price Internalizing the externalities.]TOC Social Dual
-
45Jean Walrand Porquerolles May 03
Example
x Max U1(x) + U2(y)
Subject to x + y C
Max U1(x) + U2(y)
Subject to x + y CC
yU1(x) = log(x), U2(y) = 2log(y).
User 1: Max log(x) px x = 1/pUser 2: Max 2log(y) py y = 2/p
Thus, x* = 1/p and y* = 2/p where 3/p = C.
Hence, x* = C/3, y* = 2C/3.
TOC Social Example
-
46Jean Walrand Porquerolles May 03
Example
x Max U1(x) + U2(y)
Subject to x + y C
Max U1(x) + U2(y)
Subject to x + y CC
yU1(x) = log(x), U2(y) = 2log(y).The network does not know U1 and U2. How does it choose p?Choose a price that goes up fast as x + y approaches C:
C
p(z)
z = x + y - C
TOC Social Example
-
47Jean Walrand Porquerolles May 03
TCP Implementationx* = C/3, y* = 2C/3.
Let dx(t) = a.dt bx(t)p(x(t) + y(t) C)dt [1]dy(t) = 2a.dt by(t)p(x(t) + y(t) C)dt. [2]
wherep(u) > 0 if u > 0 and p(u) = 0 if u < 0.
Then
x
y
21
[1]
[2]
C
(x*, y*)
x
yC
(x*, y*)
Dynamics:
TOC Social TCP
-
48Jean Walrand Porquerolles May 03
Why Does It Work?
11
Distance ||(x,y) (x*,y*)||decreases at each step oneach side of the boundary.
y/2C
(x*, y*)
x
Key properties: [CJ89]Push (x, y) in right directionDistributed: x does not need to know y, and vice versa
TOC SocialWhy?
-
49Jean Walrand Porquerolles May 03
ExtensionMax U1(x1) + U2(x2) + + Un(xn)
Uk(x) = wklog(x)
Subject to Ax C
Max U1(x1) + U2(x2) + + Un(xn)
Uk(x) = wklog(x)
Subject to Ax CSolution:dxk(t) = wka.dt bxk(t)pj(yj(t) Cj)dt [KMT98]
wherey(t) = Ax(t).
Thus,AIMD, with appropriate packet marking, works!
It maximizes the social welfare subject to capacity constraints!
TOC Social Extension
-
50Jean Walrand Porquerolles May 03
Virtual Queues
Imagine the constraint is utilization on link j is less than 90%
Then we can use the same packet marking scheme, butthe marking is done with virtual queues:
C
0.9C
With this scheme, we detect congestion before ithappens. The queues never really build up. [CK95, KS01]
TOC Social Virtual
-
51Jean Walrand Porquerolles May 03
Market Segmentation Basic Observation Paris Metro
TOC Segmentation
-
52Jean Walrand Porquerolles May 03
Basic Observation Why is there a first class in a plane?E.g., 20% are filled with 40% of revenues(20% of plane = 10% of seats, but 4x price ) How much space should there be in first class?
Demand d(p)
Price p
Demand
One classd(p)
Price p
Economy
1st Class
TOC Segmentation Basic
-
53Jean Walrand Porquerolles May 03
Basic Observation Of course, service should be better in first class. Service can be better because Better seats and food Less congested
TOC Segmentation Basic
-
54Jean Walrand Porquerolles May 03
Paris Metro [O98] Say that the Paris metro has two classes with identical cars. First class is more expensive than second class. As a result, first class is less congested and justifies a higher price.
Similar Examples: Palo Alto Country Club Harvard Business School
TOC Segmentation Metro
-
55Jean Walrand Porquerolles May 03
Paris Metro [O98] DiffServ:Say class k costs p(k) and has a corresponding arrival rate (k) and service rate , k = 1, 2. Average delay of class k = 1/( (k))
(1)(2)
T()
TOC Segmentation Metro
-
56Jean Walrand Porquerolles May 03
Game Theory Mini Tutorial Prisoners Dilemma: Nash Equilibrium Cournot: Nash, Stackleberg, Cooperation Mechanisms: Vickrey, VCG
TOC - Games
-
57Jean Walrand Porquerolles May 03
Prisoners Dilemma: C = Cooperate; D - Defect
One shot: Nash Equilibrium = (D, D) Repeated N times: Time N = (D, D) (D, D)N
TOC Games Prisoners
-
58Jean Walrand Porquerolles May 03
Prisoners Dilemma: Repeat forever:
Alice: I play C until you play D, then I play D forever Then (C, C) forever is a Nash equilibrium
(sub-game perfect) Note: Folk Theorem:
Play C % of the time, elseTOC Games Prisoners
-
59Jean Walrand Porquerolles May 03
CournotAlice and Bob are two firms (say Intel and AMD) that manufacture quantities x and y (resp.) of a product.
The unit price of the product is F x y.The unit production cost is H for Alice and Bob.
The profits of Alice and Bob are, respectively,
A(x, y) = (F x y)x Hx = (C x y)xB(x, y) = (F x y)y Hy = (C x y)y
TOC Games Cournot
-
60Jean Walrand Porquerolles May 03
Nash Assume Bob chooses y
Then Alice should choose x* = (C y)/2
A(x, y) = (C x y)xB(x, y) = (C x y)y
C
C
C/2
C/2x
y
y*
x*Best response functions
Equilibrium:(C/3, C/3)
Nash: (x, y) = (C/3, C/3); Rewards are (C2/9, C2/9).TOC Games Cournot
-
61Jean Walrand Porquerolles May 03
Stackelberg Assume Alice moves first and chooses x
Then Bob should choose y* = (C x)/2Consequently, Alices reward is (C x)x/2Accordingly, Alice should choose x* = C/2
A(x, y) = (C x y)xB(x, y) = (C x y)y
Stackelberg, with Alice = Leader:
(x, y) = (C/2, C/4); Rewards are (C2/8, C2/16).
TOC Games Cournot
-
62Jean Walrand Porquerolles May 03
Cooperation A(x, y) = (C x y)xB(x, y) = (C x y)y
Assume Alice and Bob agree to optimize x and yto maximize the total revenues that they then split.
In that case, A(x, y) + B(x, y) = (C z)z, z = x + yThe optimal value of z is C/2.
Cooperation
(x, y) = (C/4, C/4); Rewards are (C2/8, C2/8).
TOC Games Cournot
-
63Jean Walrand Porquerolles May 03
Summary A(x, y) = (C x y)xB(x, y) = (C x y)y
C2/8, C2/8C2/8, C2/16C2/9, C2/9A*, B*
C/4, C/4C/2, C/4C/3, C/3x*, y*
CooperationStackelbergNash
TOC Games Cournot
-
64Jean Walrand Porquerolles May 03
Mechanism Model Vickrey Auction Bidding for place in queue Bidding for CoS
TOC Games Mechanisms
-
65Jean Walrand Porquerolles May 03
Model N users bid b1 > b2 > > bN The utility of user n in position j is un = ajvn User n in position j has to pay a price pj(b1bN)
Question: How can we choose the prices so that each user n has an incentive to bid bn = vn?
(Incentive-compatible mechanism)
TOC Games Mechanisms Model
-
66Jean Walrand Porquerolles May 03
ModelDoes Bob have an incentive to bid more than Alice whenever v > bj+1? His utility increases by (aj aj+1)v. This is desirable if pj pj+1 < (aj aj+1)v whenever bj+1 < v.Thus, pj pj+1 = (aj aj+1)bj+1 is incentive-compatible.
pj = (aj aj+1)bj+1 + (aj+1 aj+2)bj+2 + VCG
= decrease in utility because Bob is in position j.
Bob: v, bid x ajvAlice: bid bj+1
a1 a2 a3 aj aj+1 aN
TOC Games Mechanisms Model
-
67Jean Walrand Porquerolles May 03
Vickrey AuctionAuction: One item is auctionedMechanism: Item sold to highest bidder who pays the second highest bid.
Here, a1 = 1, a2 = a3 = = 0.
Hence, p2 = p3 = = 0; p1 = b2.
pj = (aj aj+1)bj+1 + (aj+1 aj+2)bj+2 +
TOC Games Mechanisms Vickrey
-
68Jean Walrand Porquerolles May 03
Place in QueueAuction: Position j delay = j = - aj.
Utility of user n in position j is - vnj(i.e, v is the loss in utility per unit waiting time).
Mechanism: Ranked in order of bids.
pj = bj+1 + bj+2 + + bN= cost of additional delay imposed on others
pj = (aj aj+1)bj+1 + (aj+1 aj+2)bj+2 +
TOC Games Mechanisms Queue
-
69Jean Walrand Porquerolles May 03
Bidding for CoS [SV03]Auction: Class j utility vrj. Cj spots in class j, j = 1, , K 1, CK = .E.g., rj is the fraction of the traffic that goes through in class j.Mechanism: Fill classes in order of bids.
pj = (rj rj+1)dj+1 + (rj+1 rj+2)dj+2 + + (rK-1 rK)dK
where dk = bid of first user forced into class k.
Indeed: Here, aj = rk if position j is in class k.pj = (aj aj+1)bj+1 + (aj+1 aj+2)bj+2 +
TOC Games Mechanisms CoS
-
70Jean Walrand Porquerolles May 03
Bidding for QoS [SV03]
p4 = p3 + (0.9 0.7)b11Max. # users in class
Goodput of that class25 users bid for 4 classes of serviceUser bids: b1 > > b25
0.9, 101-1011-2021-25 0.4, 10
0.7, 10p2 = p1 + (0.4 0.2)0p3 = p2 + (0.7 0.4)b21
p1 = 00.2, Network ranksusers in decreasingbid order and fillsthe classes
Price of classes
If user n gets into class 2, hervaluation is 0.7vn and her price is p2
If user n gets into class 2, hervaluation is 0.7vn and her price is p2
Users should bid their true valuationUsers should bid their true valuation
TOC Games Mechanisms CoS
-
71Jean Walrand Porquerolles May 03
PricingOverview SLAs Packets
TOC - Pricing
-
72Jean Walrand Porquerolles May 03
Overview
UsersUsers ISP1
ISP2
ISP3
ISP4
ISPnTrafficRevenues
PricesQoS
Objectives
ISPs should have incentive to provide good services
Better services yield more revenuesSuitable redistribution of revenues among ISPs
TOC Pricing Overview
-
73Jean Walrand Porquerolles May 03
Overview
UsersUsers ISP1
ISP2
ISP3
ISP4
ISPnTrafficRevenues
PricesQoS
VERSION 1: SLAs Pricing SLAs Between ISPs Each provider chooses its prices to max. revenues
Leader-Follower or Nash; Bandwidth or QoS
VERSION 1: SLAs Pricing SLAs Between ISPs Each provider chooses its prices to max. revenues
Leader-Follower or Nash; Bandwidth or QoS
VERSION 2: PACKETS Pricing packets across ISPsEach provider chooses its prices to max. revenues
Leader-Follower, Nash, or Cooperation
VERSION 2: PACKETS Pricing packets across ISPsEach provider chooses its prices to max. revenues
Leader-Follower, Nash, or Cooperation
TOC Pricing Overview
-
74Jean Walrand Porquerolles May 03
SLAs Model Formulation Game-Theoretic Formulations Preliminary Results Conclusions
TOC Pricing SLA
-
75Jean Walrand Porquerolles May 03
ModelSource domains
Objective: maximize net utility R(x)-pS
s networkSx
Source Domain p
Vector of fixed demand (Poisson rates)
for various destinations Total rate of Accepted calls
SLA Unit SLA price(fixed for now)
TOC Pricing SLA Model
-
76Jean Walrand Porquerolles May 03
Provider Domains Objective: maximize profit
i pi Si - j pj Sj
Model
Downstream
piSiSells
pjSjBuys
Upstream Neighbors Neighbors
TOC Pricing SLA Model
-
77Jean Walrand Porquerolles May 03
Destination domainsSink trafficCharge fixed unit price p for
incoming SLAs
Model
TOC Pricing SLA Model
-
78Jean Walrand Porquerolles May 03
ModelAdditional Assumptions Type of domains:
Source, Provider, or Destination Routing: Fixed Traffic model: Loss networkErlang Fixed Point
Infinite internal capacity
TOC Pricing SLA Model TOC Pricing SLA Model
-
79Jean Walrand Porquerolles May 03
Games Formulation A: Domains Individually Optimize
Bandwidth
S T Dp1S1 S2 p2
S maximizes [over S1]: R(x) p1S1where x = q1q2
q1 = 1 E(q2, S1)q2 = 1 E(q1, S2)E(a, C) = Erlang loss for rate a, capacity C
T maximizes [over S2]: p1S1 p2S2TOC Pricing SLA Games
-
80Jean Walrand Porquerolles May 03
Games
S T Dp1S1 S2 p2
Formulation A1: Nash Game Degenerate Nash Equilibrium
(NEP)T maximizes [over S2]: p1S1 p2S2 S2 = 0
Reason:Fails to capture the incentive to provide good QoS
TOC Pricing SLA Games
-
81Jean Walrand Porquerolles May 03
GamesFormulation A2: Stackelberg Game
Leader-follower
SS1 S2 S3p1 p2 p3 T1 T2 D
T2 leads by choosing S3 T1 then follows by choosing S2 S then follows by choosing S1 S1(S2, S3) is best choice by S Knowing this, S2(S3) is best choice by T1 Knowing this, S3 is best choice by T2Example: If p1 > p2 > p3, then S1 = S2 = S3 = S1*
TOC Pricing SLA Games
-
82Jean Walrand Porquerolles May 03
GamesFormulation A2: Stackelberg Game
Shortcoming:
Not always a simple leader/follower structure
S1 D1P1
P3
1s
P2 D2
D3
S22s
Sa
Sb
Sc
P3 leadsP2 leads P1, butP1 and P2 are
also peers
TOC Pricing SLA Games
-
83Jean Walrand Porquerolles May 03
Games Formulation A3: Repeated Game
S S1 S2 S3p1 p2 p3
T1 T2 D
Strategy S: choose S1 as usual T1: S2 follows S1 T2: play S3 = S2 if S2 = S1*;
otherwise, if S2 < S1*, punish by playing S3 = 0
TOC Pricing SLA Games
-
84Jean Walrand Porquerolles May 03
Games Formulation A3: Repeated Game
S S1 S2 S3p1 p2 p3
T1 T2 D
Shortcomings: All players need to know S1*
If decreases, S1 decreases. Then S2 decreases following S1. This causes S3= 0 (punishing) Not desirable! Formulation does not reflect available information.
TOC Pricing SLA Games
-
85Jean Walrand Porquerolles May 03
GamesFormulation B. QoS as strategy Control QoS instead of bandwidth
Each provider chooses to guarantee a target level of QoS q
Amount of bandwidth it buys depends on this qand its traffic load
Concatenation of these local QoS guarantees forms the end-to-end QoS for source domains
Source domains buy bandwidth based on perceived end-to-end QoS level
TOC Pricing SLA Games
-
86Jean Walrand Porquerolles May 03
GamesFormulation B. QoS as strategy Control QoS instead of bandwidth An example
qi=1-E(ji qj, Si)S S1 S2 S3p1 p2 p3
T1 T2 D
S: maxq1 R(q1q2q3) - p1 S1(q2q3, q1)T1: maxq2 p1 S1(q2q3, q1) p2 S2(q1q3, q2)T2: maxq3 p2 S2(q1q3, q2) p3 S3(q1q2, q3)
TOC Pricing SLA Games TOC Pricing SLA Games
-
87Jean Walrand Porquerolles May 03
Preliminary ResultsSolution to the gameNumerical ExamplesProperties of the solution:
Effect of Prices
TOC Pricing SLA Results
-
88Jean Walrand Porquerolles May 03
Solution to the Game Modeled as a static game with complete
information Certain properties of the model:
Each players strategy set is q [0,1] SLA S(L,q) is a concave function of load, L and a convex function of QoS level, q n-player concave game; NEP exists (Rosen,
1965) However, it is not diagonally strict concave
TOC Pricing SLA Results Solution
-
89Jean Walrand Porquerolles May 03
Solution to the GameQuestions of interest Is the NEP unique? Can the NEP be reached by
decentralized algorithms? Is the NEP Pareto optimal? How is the end-to-end QoS at NEP
divided among domains? How do parameters, such as price,
affect the NEP?
TOC Pricing SLA Results Solution
-
90Jean Walrand Porquerolles May 03
Solution to the Game NEP in Tandem Networks NEP is unique, non-zero when prices are set
appropriately (e.g., pIN > pOUT is sufficient) Asynchronous, distributed algorithm for reaching
NEP Each domain: gradient algorithm using only
local information Each domain updates in a random order but
infinitely often No need to use traffic model; just estimate
the gradient by perturbation Can be shown to converge to the unique NEP
TOC Pricing SLA Results Solution TOC Pricing SLA Results Solution
-
91Jean Walrand Porquerolles May 03
Example 1QoS Levels
q3
q2 q1
At NEP
q*=(0.920, 0.920, 0.942)S*=(45.90, 45.90, 46.97)
S DT1 T2S1 S2 S3q1 q2 q3
p1 p2 p3
q3Best reply from T2
Trajectory ofAdaptation
NEP
Best reply from T1
q2TOC Pricing SLA Results Examples
= 50, R(x)=10xp1 = 5 , p2 = 2.5 , p3 = 1
-
92Jean Walrand Porquerolles May 03
Example 21214
1
2 3
4 5
6
7 8
9
13
15 12
7.5
88
7.57.5
64
443
34
1
11
2.5
11
10201555s
52025104s
20515103s
10155202s
101010 101s
5d4d3d2d1d
Shortest-path routing basedon hop-count
demands
TOC Pricing SLA Results Examples
-
93Jean Walrand Porquerolles May 03
Example 2
access links last-hop links
backbone links
access links
0.8410.8510.8770.8440.9240.873
q6q5q4q3 q2q1
backbone links
0.8790.9410.9150.8830.9140.826
q12q11q10q9 q8q7
last-hop links
0.9630.9730.9110.9500.9280.924
q18q17q16q15 q14q13
TOC Pricing SLA Results Examples
-
94Jean Walrand Porquerolles May 03
Effect of Prices on NEP On local and end-to-end QoSHigher ingress/egress price ratio better
local QoSUneven price ratios poor end-to-end QoS
SS1 S2 S3p1 p2 p3 T1 T2 D
25 5 1Best QoS
25 2 1Poor QoS (T2)
25 20 1Poor QoS (T1)
TOC Pricing SLA Results Prices
-
95Jean Walrand Porquerolles May 03
Effect of Prices on NEP Price as Strategy:
SS1 S2 S3p1 p2 p3 T1 T2 D
NEP of individual selection of prices is degenerate: For q2, p1 fixed, T2 has incentive to set p2 to infinity
Could formulate as Stackelberg or repeated game Same difficulties as Formulation A
Better formulation: Selection of profit margins:m1 = p1/p2 and m2 = p2/p3
TOC Pricing SLA Results Prices
-
96Jean Walrand Porquerolles May 03
Effect of Prices on NEP Profit Margin as Strategy:
SS1 S2 S3p1 p2 p3 T1 T2 D
NEP exists: equal profit margins m and equal QoS q Both m and q decrease with number of providers Numerically: These prices seems to yield best QoS Note: This result requires no internal capacity constraints
TOC Pricing SLA Results Prices TOC Pricing SLA Results Prices
-
97Jean Walrand Porquerolles May 03
Conclusions Tandem networks with profit
margin and QoS as strategy, equilibrium prices lead to stable equilibrium for QoS (need QoS faster than prices)
General Topology: Still open problemExamples show that if profit
margins are not excessive, then QoS tends to converge
TOC Pricing SLA Conclusions TOC Pricing SLA Conclusions
-
98Jean Walrand Porquerolles May 03
Packets Model Tandem Network General Topology Conclusions
TOC Pricing Packets
-
99Jean Walrand Porquerolles May 03
Model
Pricing per packet Packet contains fields [x|y] that we discuss later Ack either follows reverse path or there is a pricing
message (to convey information about the other providers)
+ p1+ p2
p1+ p2
monitor marks and process
inter-network billing info
TOC Pricing Packets Model
-
100Jean Walrand Porquerolles May 03
Tandem Network
ISPj gets revenues pjd(p) Leader-Follower:ISP1, then ISP2, Nash: ISPj maximizes pjd(p) s.t. d(p) Cj Cooperative: Subsets of ISPs charge the
same price and play Nash against the others
d(p1 + + pN)
C1p1
C2p2
C3 CNp3 pN
TOC Pricing Packets Tandem
-
101Jean Walrand Porquerolles May 03
Tandem Network: N = 2
TOC Pricing Packets Tandem
-
102Jean Walrand Porquerolles May 03
Tandem Network: N
For N 3, cooperationyields more revenuesfor all the ISPs
For N 3, cooperationyields more revenuesfor all the ISPs
TOC Pricing Packets Tandem
-
103Jean Walrand Porquerolles May 03
Tandem Network
Protocol for cooperation (nave): Using TTL, last ISP (here, N) figures out that
new connection ID X corresponds to N Last ISP informs ISP1 that X == N ISP1 selects price for X to maximize pd(p) and
marks the packets of connection ID X with the price p/N.
d(p1 + + pN)
C1p1
C2p2
C3 CNp3 pN
TOC Pricing Packets Tandem
-
104Jean Walrand Porquerolles May 03
General Topology
12
3
4
56
Route b
Route r
p(1, r)
p(3, b)p(1, b)
p(2, r)
p(5, r)
p(4, b)
p(6, r)
Nash: ISPj sets p(j, r) and p(j, b) to maximize its revenues subject to its capacity constraint
Cooperative: Maximize sum of revenues subject to all the capacity constraints; Redistribute the revenues in proportion to traffic
TOC Pricing Packets General
-
105Jean Walrand Porquerolles May 03
General Topology Case A. no constraint is active
Under cooperation prices are smaller and profits larger Case B. cooperation and Nash have the same set of
active constraints Total price for users is same under cooperation and Nash Under Nash, bottleneck providers get larger share than
under cooperation However, congestion price at bottlenecks is larger under
cooperation than under Nash, thus providing more incentive to upgrade (and drive the system to case A)
Case C. cooperation and Nash have different set of active constraints Still open
Fairness issue in the case of capacity constraintTOC Pricing Packets General TOC Pricing Packets General
-
106Jean Walrand Porquerolles May 03
Conclusions Packet pricing looks promising Bottleneck ISPs have incentive to play Nash,
others to cooperate Bottleneck has incentive to upgrade Eventually, demand is constraint and revenues
are split evenly In some networks, routing may prevent
bottleneck ISPs from overcharging Need more work on routing & protocols.
TOC Pricing Packets Conclusions TOC Pricing Packets Conclusions
-
107Jean Walrand Porquerolles May 03
Summary We believe the lack of economic incentives
explains the state of QoS To modify the incentives, one needs new
mechanisms for charging and redistributing revenues
Market segmentation is key to improved revenues
Pricing has two effects: Signal for efficient sharing of bottlenecks
(valuable traffic displaces less valuable) Extract revenues from users willingness to pay
TOC Summary
-
108Jean Walrand Porquerolles May 03
End-users / Enterprises
Content Service
Providers
Resource Service
Providers
Network Planning and Management
Technology
Economics
DataCollection
ResourcePricing
Business Relationship
DataCollection
DataCollection
Service Provider Ecosystem
Stakeholder
EconomicallyEfficientResourceAllocation
Summary Different application
requirements Enable new applications higher revenues
Different user utilities Enable market segmentation higher revenues
More flexible economic arrangements 800-service, third party
billing, Various players get fair share
of revenues Incentive for improved
services Flexible peering agreements
Revenues
Service Quality
increases improve
TOC Summary
-
109Jean Walrand Porquerolles May 03
SummaryDesirable Features of Mechanisms Everybody is a player
Not simply social maximization Designed for multi-provider and multi-business
environment Heterogeneous
Incentive-compatible Dont assume altruism
Scalable 12,000+ ASes
Flexible: Support different service types Dependable service, variable price Dependable price, variable service
Catalyst for ecology Network; Storage; Distribution; Contents;
TOC Summary
-
110Jean Walrand Porquerolles May 03
SummaryResearch Issues Necessity of economics-networking synergy
Upgrades of networks require incentives Incentives necessitate economic mechanisms Economic mechanisms require protocols
Design of protocols that are flexible Support different economic models
Suitable mechanisms Impact on industry structure
Implementation considerations Are overlays desirable? Security
TOC SummaryTOC Summary
-
111Jean Walrand Porquerolles May 03
Related Topics Pricing for Wi-Fi Access [MW03] ISP Competition [LMPRT01] Combinatorial Auctions [S90] Bidding for Nodes On Shortest Path [FPSS02]
TOC Related
-
112Jean Walrand Porquerolles May 03
References[BW98] Pio Baake, Thorsten Wichmann - On the Economics of Internet Peering -
January 1998.[CJ89] Chiu D-M and Jain R (1989). Analysis of the increase and decrease
algorithms for congestion avoidance in computer networks. Comp Networks and ISDN Sys 17: 114.
[CK95] Courcoubetis, C., G. Kesidis, A. Ridder, J. Walrand, and R. R. Weber (1995). Admission control and routing in ATM networks using inferences from measured buffer occupancy. IEEE Trans. On Communications 43, 1778 - 1784.
[CRT00] Cremer, J., P. Rey & J. Tirole ,"Connectivity in the Commercial Internet" Journal of Industrial Economics, December 2000, vol 48, n4, pp. 433-472
[CW03] Courcoubetis, C and R.R. Weber (2003). Pricing Network Services, Springer Verlag.
[FPSS02] Feigenbaum, C. H. Papadimitriou, R. Sami, S. Shenker (2002) A BGP-based Mechanism for Lowest-cost Routing,"in Proceedings of the 21st Symposium on Principles of Distributed Computing, ACM Press, New York, 2002, pp 173-182.
[FT91] Fudenberg, D. and J. Tirole (1991). Game Theory. MIT Press.[GK99] Gibbens, Kelly. Distributed connection acceptance control for a
connectionless network. [GH00] Geoff Huston Interconnection and Peering, November 2000. [JK02] Y. Jin and G. Kesidis, Feasible pricing of differentiated services for the
emerging Internet, in Proc. 40th Allerton Conference on Communications, Control and Computing, Oct. 2002.
TOC References
-
113Jean Walrand Porquerolles May 03
References[KMT98] F. P. Kelly, A.K. Maulloo D.K.H. Tan Rate control for communication
networks:shadow prices, proportional fairness and stability (1998) Journal of the Operational Research Society.
[KS01] Srisankar Kunniyur, R. Srikant Analysis and Design of an Adaptive Virtual Queue (AVQ) Algorithm for Active Queue Management Srisankar Kunniyur R. Srikant, Proceedings of SIGCOMM 2001.
[LMPRT01] J.-J. Laffont, S. Marcus, P. Rey, and J. Tirole, Internet interconnection and the off-net-cost pricing principle, 2001.
[MW03] Musacchio, J. and J. Walrand. Game-theoretic analysis of Wi-Fi Access Pricing, in preparation.
[O98] Andrew Odlyzko, Paris Metro Pricing for the Internet (1998), ACM Conference on Electronic Commerce
[PG93] Abhay K. Parekh, Robert G. Gallager: A Generalized Processor Sharing Approach to Flow Control in Integrated Services Networks: The Multiple Node Case. INFOCOM 1993: 521-530.
[S90] Vernon Smith et al., Auction Design for Composite Goods, Journal of Economic Behavior and Organization, 14 (1990) 127-129.
[SV03] Jun Shu and Pravin Varaiya, Pricing Network Services. Infocom 2003.[V61] Vickrey, W., 1961, Counterspeculation, Auctions, and Competitive Sealed
Tenders, Journal of Finance, XVI, 8-37.
TOC References
Managing QoS in a Multiprovider NetworkQoS in a Multiprovider NetworkOverviewInternet StructureInternet StructureInternet StructureInternet StructureInternet StructureNetwork StructureImplicationsDifferent ViewsInteractionsPeering AgreementsPeering AgreementsPeering AgreementsPeering AgreementsPeering Agreements - BGPPeering Agreements - BGPPeering Agreements Note: BGP is blind to utilizationApproachesWhat and Where to Control?Best EffortPer Flow ReservationPer Flow Reservation - IntServPer Flow Reservation - IntServEnd-to-End Per Flow AdmissionEnd-to-End Per Flow AdmissionEnd-to-End Per Flow AdmissionEdge-to-Edge Pipe (e.g., MPLS)Local Pipes (SLAs; e.g., DiffServ)Local Pipes (SLAs; e.g., DiffServ)Local Pipes (SLAs; e.g., DiffServ)Local Pipes (SLAs; e.g., DiffServ)Local Pipes (SLAs; e.g., DiffServ)Local Pipes (SLAs; e.g., DiffServ)Local Pipes (SLAs; e.g., DiffServ)Local Pipes (SLAs; e.g., DiffServ)Some LessonsComparisonComparisonWhy not much QoS?Social Welfare and Packet MarkingPrimalDualExampleExampleTCP ImplementationWhy Does It Work?ExtensionVirtual Queues