layering as optimization decomposition: a mathematical theory of
TRANSCRIPT
![Page 1: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/1.jpg)
Layering As Optimization Decomposition:
A Mathematical Theory of Network Architectures
Mung Chiang
Electrical Engineering Department, Princeton
Steven H. Low
Computer Science and Electrical Engineering Departments, Caltech
A. Robert Calderbank
Electrical Engineering and Mathematics Departments, Princeton
IEEE ISIT Tutorial
July 9, 2006
![Page 2: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/2.jpg)
Schedule of the Tutorial
2:00 – 2:30pm Overview (Chiang)
2:30 – 3:00pm TCP: reverse and forward engineering (Low)
3:00 – 3:20pm MAC: reverse and forward engineering (Chiang)
3:20 – 3:35pm Decomposition theory and alternative decompositions
(Chiang)
3:35 – 3:45pm Break
3:45 – 4:00pm Case 1: Joint congestion control and coding
(Calderbank)
4:00 – 4:15pm Case 2: Joint congestion control, routing, and
scheduling (Chiang)
4:15 – 4:30pm Case 3: TCP/IP interactions (Low)
4:30 – 4:50pm Future research challenges and Summary (30 open
issues, 20 methodologies, 10 key messages) (Chiang)
4:50 – 5:00pm Question and Answers
![Page 3: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/3.jpg)
Part I
Overview
![Page 4: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/4.jpg)
Nature of the Tutorial
M. Chiang, S. H. Low, A. R. Calderbank, and J. C. Doyle, “Layering as
optimization decomposition: A mathematical theory of network
architectures” Proceedings of IEEE, December 2006.
• Give an overview of the topic. Details in various papers
• Not exhaustive survey. Highlight the key ideas and challenges
• Biased presentation. Focus on work by us
This is an appetizer. The beef in the papers
![Page 5: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/5.jpg)
Outline
• Background: Holistic view on layered architecture
• Background: NUM and G.NUM
• Horizontal Decompositions
• Vertical Decompositions
• Alternative Decompositions
• Key Messages
• Key Methodologies
• Open Issues
![Page 6: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/6.jpg)
Acknowledgement
Collaborators:
Stephen Boyd, Lijun Chen, Dah Ming Chiu, Neil Gershenfeld, Andrea
Goldsmith, Prashanth Hande, Jiayue He, Sanjay Hegde, Maryam Fazel,
Cheng Jin, Koushik Kar, Frank Kelly, P. R. Kumar, Jang-Won Lee,
Ruby Lee, Lun Li, Ying Li, Xiaojun Lin, Jiaping Liu, Zhen Liu, Asuman
Ozdaglar, Daniel Palomar, Pablo Parrilo, Jennifer Rexford, Devavrat
Shah, Ness Shroff, R. Srikant, Chee Wei Tan, Ao Tang, Jiantao Wang,
Xin Wang, David Wei, Bartek Wydrowski, Lin Xiao, Edmund Yeh, and
Junshan Zhang
Funding Agencies:
NSF grants CCF-0440043, CCF-0448012, CNS-0417607,
CNS-0427677, CNS-0430487, CNS-0519880, DARPA grant
HR0011-06-1-0008, and Cisco grant GH072605
![Page 7: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/7.jpg)
Layered Network Architecture
Application
Presentation
Session
Transport
Network
Link
Physical
Important foundation for data networking
Ad hoc design historically (within and across layers)
![Page 8: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/8.jpg)
Rethinking Layering
How to, and how not to, layer? A question on architecture
Functionality allocation: who does what and how to connect them?
But want answers to be rigorous, quantitative, simple, and relevant
• How to modularize (and connect)?
• How to distribute (and connect)?
• How to search in the design space of alternative architectures?
• How to quantify the benefits of better
codes/modulation/schedule/routes... for network applications?
A common language to rethink these issues?
![Page 9: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/9.jpg)
The Goal
A Mathematical Theory of Network Architectures
• Particular focus on the architectures of layering and distributed control
• There are also boundaries to the use of mathematical approach to the
economics, psychology, and engineering of network architectures
• But certainly provides rigorous approaches on why protocols work,
when it will not work, and how to make it work better
• Also provides conceptually clear understanding on the opportunities
and risks of cross layer design
![Page 10: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/10.jpg)
Layering As Optimization Decomposition
The first unifying view and systematic approach
Network: Generalized NUM
Layering architecture: Decomposition scheme
Layers: Decomposed subproblems
Interfaces: Functions of primal or dual variables
Horizontal and vertical decompositions through
• implicit message passing (e.g., queuing delay, SIR)
• explicit message passing (local or global)
3 Steps: G.NUM ⇒ A solution architecture ⇒ Alternative architectures
![Page 11: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/11.jpg)
Network Utility Maximization
Basic NUM (KellyMaulloTan98):
maximizeP
s Us(xs)
subject to Rx ¹ c
x º 0
Generalized NUM (one possibility shown here) (Chiang05a):
maximizeP
s Us(xs, Pe,s) +P
j Vj(wj)
subject to Rx ¹ c(w,Pe)
x ∈ C1(Pe)
x ∈ C2(F) or x ∈ Π
R ∈ RF ∈ Fw ∈ W
![Page 12: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/12.jpg)
GNUM
• Objective function: What the end-users and network provider care
about (can be coupled, eg, one utility function for the whole network)
• Constraint set: Physical and economic limitations
• Variables: Under the control of this design
• Constants: Beyond the control of this design
![Page 13: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/13.jpg)
Two Cornerstones for Conceptual Simplicity
Networks as optimizers
Reverse engineering mentality: give me the solution (an existing
protocol), I’ll find the underlying problem implicitly being solved
• Why care about the problem if there’s already a solution?
• It leads to simple, rigorous understanding for systematic design
Layering as decomposition
1. Analytic foundation for network architecture
2. Common language for thinking and comparing
3. Methodologies, analytic tools
![Page 14: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/14.jpg)
Layering as Optimization Decomposition
What’s so unique about this particular framework for cross-layer design?
• Network as optimizer
• End-user application utilities as the driver
• Performance benchmark without any layering
• Unified approach to cross-layer design (it simplifies our understanding
about network architecture)
• Separation theorem among modules
• Systematic exploration of architectural alternatives
Not every cross-layer paper is ‘layering as optimization decomposition’
![Page 15: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/15.jpg)
Utility
Which utility? (function of rate, useful information, delay, energy...)
1. Reverse engineering: TCP maximizes utilities
2a. Behavioral model: user satisfaction
2b. Traffic model: traffic elasticity
3a. Economics: resource allocation efficiency
3b. Economics: different utility functions lead to different fairness
Three choices: Weighted sum, Pareto optimality, Uncooperative game
• Goal: Distributed and modularized algorithm converging to globally
and jointly optimum resource allocation
• Limitations to be discussed at the end
![Page 16: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/16.jpg)
Layers
Restriction: we focus on resource allocation functionalities rather than
semantics functionalities
• TCP: congestion control
Different meanings:
• Routing: RIP/OSPF, BGP, wireless routing, optical routing,
dynamic/static, single-path/multi-path, multicommodity flow routing...
• MAC: scheduling or contention-based
• PHY: power control, coding, modulation, antenna signal processing...
Insights on both:
• What each layer can do (Optimization variables)
• What each layer can see (Constants, Other subproblems’ variables)
![Page 17: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/17.jpg)
Connections With Mathematics
• Convex and nonconvex optimization
• Decomposition and distributed algorithm
• Game theory, General market equilibrium theory
• Algebraic geometry (nonconvex formulations)
• Differential topology (heterogeneous protocols)
![Page 18: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/18.jpg)
Adoption By Industry
Industry adoption of Layering As Optimization Decomposition:
• Internet resource allocation: TCP FAST (Caltech)
• Protocol stack design: Internet 0 (MIT)
• Broadband access: FAST Copper (Princeton, Stanford, Fraser)
This tutorial is mainly about the underlying common language and
methodologies
![Page 19: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/19.jpg)
Network Utility Maximization
BNUM (KellyMaulloTan98):
maximizeP
s Us(xs)
subject to Rx ¹ c
x º 0
GNUM (one possibility shown here) (Chiang05a):
maximizeP
s Us(xs, Pe,s) +P
j Vj(wj)
subject to Rx ¹ c(w,Pe)
x ∈ C1(Pe)
x ∈ C2(F) or x ∈ Π
R ∈ RF ∈ Fw ∈ W
![Page 20: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/20.jpg)
Dual-based Distributed Algorithm
BNUM with concave smooth utility functions:
Convex optimization (Monotropic Programming) with zero duality gap
Lagrangian decomposition:
L(x, λ) =X
s
Us(xs) +X
l
λl
0@cl −
X
s:l∈L(s)
xs
1A
=X
s
24Us(xs)−
0@ X
l∈L(s)
λl
1Axs
35+
X
l
clλl
=X
s
Ls(xs, λs) +X
l
clλl
Dual problem:
minimize g(λ) = L(x∗(λ), λ)
subject to λ º 0
![Page 21: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/21.jpg)
Dual-based Distributed Algorithm
Source algorithm:
x∗s(λs) = argmax [Us(xs)− λsxs] , ∀s
• Selfish net utility maximization locally at source s
Link algorithm (gradient or subgradient based):
λl(t + 1) =
24λl(t)− α(t)
0@cl −
X
s:l∈L(s)
xs(λs(t))
1A35
+
, ∀l
• Balancing supply and demand through pricing
Certain choices of step sizes α(t) of distributed algorithm guarantee
convergence to globally optimal (x∗, λ∗)
![Page 22: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/22.jpg)
Primal-Dual
Different meanings:
• Primal-dual interior-point algorithm
• Primal-dual solution
• Primal or dual driven control
• Primal or dual decomposition
Coupling in constraints (easy: flow constraint, hard: SIR feasibility)
Coupling in objective (easy: additive form, hard: min max operations)
![Page 23: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/23.jpg)
Primal Decomposition
Simple example:
x + y + z + w ≤ c
Decomposed into:
x + y ≤ α
z + w ≤ c− α
New variable α updated by various methods
Interpretation: Direct resource allocation (not pricing-based control)
Engineering implications: Adaptive slicing (GENI)
Pricing feedback: dual decomposition
Adaptive slicing: primal decomposition
![Page 24: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/24.jpg)
Horizontal Decompositions
Reverse engineering:
• Layer 4 TCP congestion control: Basic NUM (LowLapsley99,
RobertsMassoulie99, MoWalrand00, YaicheMazumdarRosenberg00,
KunniyurSrikant02, LaAnatharam02, LowPaganiniDoyle02, Low03,
Srikant04...)
• Layer 4 TCP heterogeneous protocol: Nonconvex equilibrium problem
(TangWangLowChiang05)
• Layer 3 IP inter-AS routing: Stable Paths Problem
(GriffinSheperdWilfong02)
• Layer 2 MAC backoff contention resolution: Non-cooperative Game
(LeeChiangCalderbank06a)
Forward engineering for horizontal decompositions also carried out
recently
![Page 25: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/25.jpg)
Vertical Decompositions
A partial list of work along this line:
• Jointly optimal congestion control and adaptive coding or power
control (Chiang05a, LeeChiangCalderbank06b)
• Jointly optimal congestion and contention control
(KarSarkarTassiulas04, ChenLowDoyle05, WangKar05, YuenMarbach05,
ZhengZhang06, LeeChiangCalderbank06c)
• Jointly optimal congestion control and scheduling (ErilymazSrikant05)
• Jointly optimal routing and scheduling (KodialamNandagopal03)
• Jointly optimal routing and power control (XiaoJohanssonBoyd04,
NeelyModianoRohrs05)
• Jointly optimal congestion control, routing, and scheduling
(LinShroff05, ChenLowChiangDoyle06)
![Page 26: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/26.jpg)
Vertical Decompositions
• Jointly optimal routing, scheduling, and power control
(CruzSanthanam03, XiYeh06)
• Jointly optimal routing, resource allocation, and source coding
(YuYuan05)
• TCP/IP interactions (WangLiLowDoyle05, HeChiangRexford06) and
jointly optimal congestion control and routing (KellyVoice05,
Hanetal05)
• Network lifetime maximization (NamaChiangMandayam06)
• Application adaptation and congestion control/resource allocation
(ChangLiu04, HuangLiChiangKatsaggelos06)
Apology, Apology, Apology for any missing reference
![Page 27: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/27.jpg)
Vertical Decompositions
• Specific designs not important
• Common language and key messages methodologies important
Goal: Shrink, not grow knowledge tree on cross-layer design
![Page 28: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/28.jpg)
Alternative Decompositions
Many ways to decompose:
• Primal and dual decomposition
• Partial decomposition
• Multi-level decomposition
Master Problem
Subproblem 1...
Secondary Master Problem
prices / resources
First LevelDecomposition
Second LevelDecomposition
Subproblem
Subproblem N
prices / resources
Lead to alternative architectures (PalomarChiang06) with different
• Communication overhead
• Computation distribution
• Convergence behavior
![Page 29: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/29.jpg)
Alternative Decompositions
XAlternative
problem representations
Different algorithms
Engineering implications
X X...
...
...
Systematically explore the space of alternative decompositions
![Page 30: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/30.jpg)
Key Messages
• Protocols in layers 2,3,4 can be reverse engineered. Reverse
engineering in turn leads to better design in a rigorous manner.
• There is a unifying approach to cross-layer design, illustrating both
opportunities and risks.
• Loose coupling through “layering price” can be optimal and robust,
and congestion price (or queuing delay, or buffer occupancy) is often
the right “layering price” for stability and optimality, with important
exceptions as well.
• User-generated pricing following end-to-end principle
• There are many alternatives in decompositions, leading to different
divisions of tasks across layers and even different time-scales of
interactions.
• Convexity of the generalized NUM is the key to devising a globally
optimal solution.
• Decomposability of the generalized NUM is the key to devising a
distributed solution.
![Page 31: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/31.jpg)
Key Methodologies
• Dual decomposition for linear coupling constraints.
• Consistency pricing for coupled objective functions.
• Descent lemma for proof of convergence of dual-based distributed
subgradient algorithm.
• Stability proof through Lyapunov function construction, singular
perturbation theory, and passivity argument.
• Log change of variables to turn multiplicative coupling into linear
coupling, and to turn nonconvex constraints to convex ones.
• Sufficient conditions on curvature of utility functions for it to remain
concave after a log change of variables.
• Construction of conflict graph, contention matrix, and transmission
modes in contention based MAC design.
• Maximum differential congestion pricing for node-based back-pressure
scheduling (part of the connections between distributed convex
optimization and stochastic control).
![Page 32: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/32.jpg)
Future Research Issues
• Technical: Global stability under delay...
• Modeling: routing in ad hoc network, ARQ, MIMO...
• Time issues
• Why deterministic fluid model?
Shannon 1948: remove finite blocklength, Law of Large Numbers kicks
in (later finite codewords come back...)
Kelly 1998: remove coupled queuing dynamics, optimization and
decomposition view kicks in (later stochastics come back...)
• What if it’s not convex optimization?
Rockafellar 1993: Convexity is the watershed between easy and hard
(what if it’s hard?)
• Is performance the only optimization objective?
![Page 33: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/33.jpg)
Future Research: Time Issues
• Rate of convergence
• Timescale separation
• Transient behavior bounding
• Utility as a function of latency
• Utility as a function of transient rate allocations
![Page 34: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/34.jpg)
Future Research: Stochastic Issues
Fill the table with 3 stars in all entries:
Union of Stochastic Network and Network Optimization
Stability or Average Outage Fairness
Validation Performance Performance
Session Level ?? ? ?
Packet Level ? ?
Channel Level ?? ?
Topology Level
Table 1: State-of-the-art in Stochastic Network Utility Maximization.
With a good layering architecture:
• Stochastic doesn’t hurt
• Stochastic may help
![Page 35: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/35.jpg)
Future Research: Nonconvexity Issues
• Nonconcave utility (eg, real-time applications)
• Nonconvex constraints (eg, power control in low SIR)
• Integer constraints (eg, single-path routing)
• Exponentially long description length (eg, scheduling)
Convexity not invariant under embedding in higher dimensional space or
nonlinear change of variable
• Sum-of-squares method (Stengle73, Parrilo03)
• Geometric programming (DuffinPetersonZener67, Chiang05b)
From optimal/complicated to suboptimal/simple modules (LinShroff05)
![Page 36: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/36.jpg)
Future Research: Network X-ities Issues
From Bit to Utility to Control and Management
Over-optimized? Optimizing for what?
• Evolvability
• Scalability
• Diagnosability
Pareto-optimal tradeoff between Performance and Network X-ities
From Forward Engineering to Reverse Engineering to
• Design for Optimizability
![Page 37: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/37.jpg)
More later
See lists of
• 30 open issues
• 20 methodologies
• 10 key messages
at the end of the tutorial
![Page 38: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/38.jpg)
New Mentalities
Layering As Optimization Decomposition, but move away from:
• One architecture fits all
• Deterministic fluids
• Asymptotic convergence
• Optimality
• Optimization
Think about “right” decomposition in the “right” way
![Page 39: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/39.jpg)
Contacts
www.princeton.edu/∼chiangm
netlab.caltech.edu
![Page 40: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/40.jpg)
Part II
TCP Congestion Control: Reverse and Forward Engineering
![Page 41: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/41.jpg)
Some References
• F. P. Kelly, A. Maulloo, and D. Tan, “Rate control for communication
networks: shadow prices, proportional fairness and stability,” Journal of
Operations Research Society, vol. 49, no. 3, pp.237-252, March 1998
• S. H. Low, “A duality model of TCP and queue management
algorithms,” IEEE/ACM Transactions on Networking, vol. 11, no. 4,
pp. 525-536, August 2003
• R. Srikant, The Mathematics of Internet Congestion Control,
Birkhauser, 2004
• C. Jin, D. X. Wei, and S. H. Low, “FAST TCP: Motivation,
architecture, algorithms, and performance”, Proc. IEEE INFOCOM,
March 2004
• A. Tang, J. Wang, S. H. Low, and M. Chiang, “Network equilibrium
of heterogeneous congestion control protocols,” Proc. IEEE
INFOCOM, March 2005
![Page 42: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/42.jpg)
TCP Congestion Control
Reverse engineering: B.NUMForward engineering:
FASTHeterogeneous protocols
![Page 43: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/43.jpg)
c1 c2
Congestion control
Network: links l with capacity cl
Sources s: L(s) - links used by source sTCP: dynamically adapts xs to congestion to ensure
x1
x2x3
121 cxx ≤+ 231 cxx ≤+
lcx liLli
i links allfor )(:
≤∑∈
![Page 44: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/44.jpg)
c1 c2
Congestion control
Challenge: available info must be end-to-endImplicit congestion feedback
Loss probability: likelihood of a packet being delivered correctlyRound-trip time: time it takes for a packet to reach its destination and for its ack to return to the sender
Explicit congestion feedback: marks, rates x1
x2x3
121 cxx ≤+ 231 cxx ≤+
![Page 45: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/45.jpg)
TCP & AQM
xi(t)
pl(t)
TCP: RenoVegasHSTCPFASTSTCP
AQM:DropTailREDREM/PIAVQ
![Page 46: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/46.jpg)
Reverse engineering
Protocol (Reno, Vegas, RED, REM/PI…)
EquilibriumPerformance
Throughput, loss, delayFairnessUtility
DynamicsLocal stabilityGlobal stability
))( ),(( )1())( ),(( )1(txtpGtptxtpFtx
=+=+
![Page 47: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/47.jpg)
Network model
F1
FN
G1
GL
R
RT
TCP Network AQM
x y
q p
))( ),(( )1())( ),(( )1(
tRxtpGtptxtpRFtx T
=+=+ Reno, Vegas
DT, RED, …
liRli link uses source if 1= IP routing
![Page 48: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/48.jpg)
for every RTT
{ W += 1 }
for every loss
{ W := W/2 }
(AI)
(MD)
Reno:Jacobson
1989
⎟⎠
⎞⎜⎝
⎛=+
−=+
∑
∑
ililill
llli
i
ii
tptxRGtp
tpRxT
tx
)(),()1(
)(2
1)1(2
2
AI
MD
TailDrop
Network model: example
![Page 49: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/49.jpg)
FAST:Jin, Wei, Low
2004α W
RTTbaseRTT :W +=
periodically
{
}
⎟⎠
⎞⎜⎝
⎛−+=+
⎟⎠
⎞⎜⎝
⎛−+=+
∑
∑
ilili
lll
llliii
i
iii
ctxRc
tptp
tpRtxT
txtx
)(1)()1(
)()()()1( αγ
Network model: example
![Page 50: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/50.jpg)
Duality model of TCP/AQMTCP/AQM
Equilibrium (x*,p*) primal-dual optimal:
F determines utility function UG guarantees complementary slackness
p* are Lagrange multipliers
) ,( ) ,(
***
***
RxpGpxpRFx T
=
=
cRxxU iix≤∑
≥ subject to )( max
0
Uniqueness of equilibriumx* is unique when U is strictly concavep* is unique when R has full row rank
Kelly, Maloo, Tan 1998Low, Lapsley 1999
![Page 51: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/51.jpg)
Duality model of TCP/AQMTCP/AQM
Equilibrium (x*,p*) primal-dual optimal:
F determines utility function UG guarantees complementary slackness
p* are Lagrange multipliers
) ,( ) ,(
***
***
RxpGpxpRFx T
=
=
cRxxU iix≤∑
≥ subject to )( max
0
Kelly, Maloo, Tan 1998Low, Lapsley 1999
The underlying concave program also leads to simple dynamic behavior
![Page 52: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/52.jpg)
Duality model of TCP/AQM
Equilibrium (x*,p*) primal-dual optimal: cRxxU iix
≤∑≥
subject to )( max0
Mo & Walrand 2000:
⎪⎩
⎪⎨⎧
≠−
== −− 1 if )1(
1 if log)( 11 αα
αα
i
iii x
xxU
α = 1 : Vegas, FAST, STCP α = 1.2: HSTCPα = 2 : Renoα = : XCP (single link only) ∞
![Page 53: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/53.jpg)
222
2
3
33
)1(4)1 )(
2 ππββπττρ
−+≤+⋅
-(NcN
c
Theorem (Low et al, Infocom’02)
Reno/RED is locally stable if
F1
FN
Stability: Reno/RED
G1
GL
Rf(s)
Rb’(s)
TCP Network AQM
x y
q p
TCP:Small τSmall cLarge N
RED:Small ρLarge delay
![Page 54: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/54.jpg)
Stability: scalable control
F1
FN
G1
GL
Rf(s)
Rb’(s)
TCP Network AQM
x y
q p
( )lll
l ctyc
tp −= )(1 )(&)(
)(tq
mii
iii
i
extx τα
−
=
Theorem (Paganini, Doyle, L, CDC’01)
Provided R is full rank, feedback loop is locally stable for arbitrary delay and capacity
![Page 55: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/55.jpg)
Stability: FAST
⎟⎟⎠
⎞⎜⎜⎝
⎛−⋅=
)()(1)(
tutqtw
i
iii κ&
F1
FN
G1
GL
Rf(s)
Rb’(s)
TCP Network AQM
x y
q p
( )lll
l ctyc
tp −= )(1 )(&
ApplicationStabilize TCP with current routersQueueing delay as congestion measure has right scalingIncremental deployment with ECN
![Page 56: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/56.jpg)
Some implicationsEquilibrium
Always exists, unique if R is full rankBandwidth allocation independent of AQM or arrivalCan predict macroscopic behavior of large scale networks
Counter-intuitive throughput behaviorFair allocation is not always inefficient Increasing link capacities do not always raise aggregate throughput
[Tang, Wang, Low, ToN 2006]
FAST TCPDesign, analysis, experiments
[Jin, Wei, Low, ToN 2007]
![Page 57: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/57.jpg)
Duality modelHistorically
Packet level implemented firstFlow level understood as after-thoughtBut flow level design determines
performance, fairness, stability
Now: can forward engineerSophisticated theory on equilibrium & stability (optimization+control)Given (application) utility functions, can design provably scalable TCP algorithms
![Page 58: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/58.jpg)
Packet level
ACK: W W + 1/W
Loss: W W – 0.5W
RenoAIMD(1, 0.5)
ACK: W W + a(w)/W
Loss: W W – b(w)W
HSTCPAIMD(a(w), b(w))
ACK: W W + 0.01
Loss: W W – 0.125W
STCPMIMD(a, b)
α RTT
baseRTT W W :RTT +⋅←FAST
![Page 59: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/59.jpg)
Flow level: Reno, HSTCP, STCP, FAST
Different gain κ and utility UiThey determine equilibrium and stability
Different congestion measure piLoss probability (Reno, HSTCP, STCP)Queueing delay (Vegas, FAST)
Common flow level dynamics!
⎟⎟⎠
⎞⎜⎜⎝
⎛−⋅=
)()(1 )( )( ' tU
tpttwi
ii κ&
windowadjustment
controlgain
flow levelgoal=
![Page 60: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/60.jpg)
Flow level: Reno, HSTCP, STCP, FAST
Similar flow level equilibrium
FAST
1 STCP
1 HSTCP
1 Reno
84.0
5.0
ii
iii
iii
iii
px
pTx
pTx
pTx
α
α
α
α
=
⋅=
⋅=
⋅=
α = 1.225 (Reno), 0.120 (HSTCP), 0.075 (STCP)
pkts/sec
![Page 61: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/61.jpg)
FAST Architecture
Burstiness Control
Window Control
TCP Protocol Processing
DataControl
Estimation
RTT timescaleLoss recovery
<RTT timescale
![Page 62: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/62.jpg)
FAST Architecture
Burstiness Control
Window Control
TCP Protocol Processing
DataControl
Estimation
Each componentdesigned independentlyupgraded asynchronously
![Page 63: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/63.jpg)
FAST Architecture
Burstiness Control
Window Control
TCP Protocol Processing
DataControl
Estimation
Each componentdesigned independentlyupgraded asynchronously
WindowControl
![Page 64: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/64.jpg)
( )
lii
i
iiii
ctqd
twtxtqtwtw
=+
−+=+
∑ )()( : clocking-self
)()()()1( : updatewindow αγ
Window control algorithm
![Page 65: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/65.jpg)
Window control algorithm
Theorem (CDC04, Infocom05, IMA06)
Utility function: αi log xi (proportional fairness)
Locally stable in networks if feedback delays are homogeneous (e.g. zero)
Global exponential convergence over a single link
( )
lii
i
iiii
ctqd
twtxtqtwtw
=+
−+=+
∑ )()( : clocking-self
)()()()1( : updatewindow αγ
![Page 66: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/66.jpg)
Network
(Sylvain Ravot, caltech/CERN)
![Page 67: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/67.jpg)
FAST TCPutil: 95%
Linux TCPutil: 19%
1Gbps path; 180 ms RTT; 1 flowJin, Wei, Ravot, etc (Caltech, Nov 02)
![Page 68: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/68.jpg)
Abilene Test
OC48
OC192
(Yang Xia, Harvey Newman, Caltech)
Periodic lossesevery 10mins
![Page 69: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/69.jpg)
(Yang Xia, Harvey Newman, Caltech)
Periodic lossesevery 10mins
![Page 70: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/70.jpg)
Emulated WAN emulates different operating conditionsMeasure application throughput (iperf) with and without Aria 1000Throughput improvements are similar with Windows 2003 and Linux
Benchmark testbedsender receiver
emulated WAN
FastSoftAria 1000TM emulated
WAN
![Page 71: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/71.jpg)
Linux (tuned)Linux (default)
+ Aria 1000
WAN speed: 45Mbps
0.00% 0.01% 0.10%1.00%
LA-HK (200ms)
LA-NY (100ms)LA-SF (50ms)
05
1015202530
35
40
45
Thro
ughp
ut (M
bps)
Loss Rate
Up to 22x improvement on T3@ 0.1% loss rate
0.00% 0.01% 0.10% 1.00%
LA-HK (200ms)LA-NY (100ms)
LA-SF (50ms)0
10
20
30
40
Thro
ughp
ut (M
bps)
Loss Rate
![Page 72: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/72.jpg)
Linux (tuned) Linux (default)+ Aria 1000
0.00% 0.01% 0.10% 1.00%
LA-HK (200ms)LA-NY (100ms)
LA-SF (50ms)0
20406080
100120140
Thro
ughp
ut (M
bps)
Loss Rate 0.00% 0.01% 0.10% 1.00%
LA-HK (200ms)
LA-NY (100ms)LA-SF (50ms)
0
20
40
60
80
100
120
140
Thro
ughp
ut (M
bps)
Loss Rate
WAN speed: 155Mbps
Up to 17x improvement on OC3.01% loss rate
![Page 73: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/73.jpg)
The world is heterogeneous…
Linux 2.6.13 allows users to choose congestion control algorithmsMany protocol proposals
Loss-based: Reno and a large number of variantsDelay-based: CARD (1989), DUAL (1992), Vegas (1995), FAST (2004), …ECN: RED (1993), REM (2001), PI (2002), AVQ (2003), …Explicit feedback: MaxNet (2002), XCP (2002), RCP (2005), …
![Page 74: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/74.jpg)
Some implications
homogeneous heterogeneous
equilibrium unique non-unique
bandwidthallocation on AQM
independent dependent
bandwidthallocationon arrival
independent dependent
![Page 75: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/75.jpg)
F1
FN
G1
GL
R
RT
TCP Network AQM
x y
q p
Homogeneous protocol
( ) ⎟⎠
⎞⎜⎝
⎛=+
⎟⎠
⎞⎜⎝
⎛=+
∑
∑
)( ,)( )1(
)( ,)( )1(
txtpmRFtx
txtpRFtx
ji
ll
jlli
ji
ji
il
lliii
same pricefor all sources
![Page 76: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/76.jpg)
F1
FN
G1
GL
R
RT
TCP Network AQM
x y
q p
Heterogeneous protocol
( ) ⎟⎠
⎞⎜⎝
⎛=+
⎟⎠
⎞⎜⎝
⎛=+
∑
∑
)( ,)( )1(
)( ,)( )1(
txtpmRFtx
txtpRFtx
ji
ll
jlli
ji
ji
il
lliii
heterogeneousprices for
type j sources
![Page 77: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/77.jpg)
Heterogeneous protocols
Equilibrium: p that satisfies
∑
∑
⎩⎨⎧
>=≤
=
⎟⎠
⎞⎜⎝
⎛=
i,j ll
lji
jlil
ll
jlli
ji
ji
pcc
pxRpy
pmRfpx
0 if
)( : )(
)( )(
Dynamic: dual algorithm
( )llll
ll
jlli
ji
ji
ctpyp
tpmRftpx
−=
⎟⎠
⎞⎜⎝
⎛= ∑
))((
))(( ))((
γ&
![Page 78: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/78.jpg)
Existence
TheoremEquilibrium p exists, despite lack of
underlying utility maximization
Generally non-uniqueThere are networks with unique bottleneck set but infinitely many equilibriaThere are networks with multiple bottleneck set each with a unique (but distinct) equilibrium
![Page 79: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/79.jpg)
Regular networks
DefinitionA regular network is a tuple (R, c, m, U) for
which all equilibria p are locally unique, i.e.,
TheoremAlmost all networks are regularA regular network has finitely many and odd number of equilibria (e.g. 1)
Proof: Sard’s Theorem and Poincare-HopfIndex Theorem
0 )(det : )(det ≠∂∂
= ppypJ
![Page 80: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/80.jpg)
Global uniqueness
CorollaryIf price mapping functions ml
j are linear and link-independent, then equilibrium is globally unique
e.g. a network of RED routers almost always has globally unique equilibrium
TheoremIf Degree of heterogeneity is small, then equilibrium is globally unique
0any for ]2,[
0any for ]2,[/1
/1
>∈
>∈jjLjj
l
llL
lj
l
aaam
aaam&
&
![Page 81: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/81.jpg)
Local stability:`uniqueness’ stability
TheoremIf Degree of heterogeneity is small, then the unique equilibrium p is locally stable
0any for ]2,[
0any for ]2,[/1
/1
>∈
>∈jjLjj
l
llL
lj
l
aaam
aaam&
&
p(t)pp δγδ )( *J=&Linearized dual algorithm:
Equilibrium p is locally stable if
( ) 0 )( Re <pJλ
![Page 82: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/82.jpg)
Local stability:`converse’
Theorem
If all equilibria p are locally stable, then it is globally unique
LpI )1( )( −=Proof idea:
For all equilibrium p:Index theorem:
L
ppI )1( )(
eq−=∑
![Page 83: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/83.jpg)
FAST throughput
without slow timescale control with slow timescale control
0 200 400 600 800 1000 1200 1400 1600 18000
1
2
3
4
5
6
7
time(sec)
thro
ug
hp
ut
(pkt
/ m
s)
FASTReno
Forward engineering: ns2 simulation
![Page 84: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/84.jpg)
Part III
Medium Access Control: Reverse and Forward Engineering
![Page 85: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/85.jpg)
Some References
• J. W. Lee, M. Chiang, and R. A. Calderbank, “Utility-optimal medium
access control: reverse and forward engineering,” Proc. IEEE
INFOCOM, April 2006
• X. Wang and K. Kar, “Cross-layer rate control for end-to-end
proportional fairness in wireless networks with ranom access”, IEEE
Journal of Selected Areas in Communications, vol. 24, no. 8, August
2006
• J. W. Lee, M. Chiang, and A. R. Calderbank, “Jointly optimal
congestion and contention control”, IEEE Communication Letters, vol.
10, no. 3, pp. 216-218, March 2006
![Page 86: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/86.jpg)
Utility Approach
Two approaches of understanding and designing random access MAC:
• Queuing theoretic: stochastic stability
• Optimization theoretic: utility optimality
Eventually, we want both in a unifying framework
• A lot of results on the first
• This part of the tutorial is about the second
![Page 87: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/87.jpg)
Reverse Engineering
What are heuristics-based protocol designs implicitly solving?
• Layer 4 TCP congestion control: Basic NUM
• Layer 3 IP inter-AS routing: Stable Paths Problem
• Layer 2 MAC backoff contention resolution: Non-cooperative Game
Focus of this part of the talk
![Page 88: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/88.jpg)
Reverse Engineering
Different from imposing a game model:
• MacKenzie Wicker 2003
• Jin Kesidis 2004
• Marbach Yuen 2005
• Altman et. al. 2005
![Page 89: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/89.jpg)
Network Topology
3 4A B C D E
G
1 2
F
I H
5
6
d d d d
d
d
Directed graph G(V, E)
LIto(l): set of links whose transmissions interfere with receiver of link l
LIfrom(l): set of links whose transmissions get interferred by
transmission on link l
![Page 90: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/90.jpg)
Exponential Backoff MAC
MAC:
• Contention-free: centralized scheduling
• Contention-based: distributed random access (contention avoidance
and resolution)
Contention resolution through exponential backoff
Binary Exponential Backoff (BEB) in IEEE 802.11 standard
Persistence probabilistic model:
• Each logical link l transmits with persistence probability pl
• Successful transmission: pl = pl,max (i.e., 1/W minl )
• Collided transmission: pl = max{βlpl, pminl }, βl ∈ (0, 1)
![Page 91: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/91.jpg)
It’s A Game
TCP/AQM: social welfare (network utility) cooperative maximization
EB-MAC: non-cooperative game
• Coupled utility (due to collision)
• Inadequate feedback
Game: [E,Q
l∈E Al, {Ul}l∈E ] with Al = {pl | pminl ≤ pl ≤ pmax
l }
• S(p) = plQ
n∈LIto(l)(1− pn): probability of transmission success
• F (p) = pl(1−Q
n∈LIto(l)(1− pn)): probability of collision
![Page 92: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/92.jpg)
Reverse-Engineering Utility Function
Theorem: Ul(p) = R(pl)S(p)− C(pl)F (p), with
R(pl)def= pl(
12pmax
l − 13pl) (reward for transmission success)
C(pl)def= 1
3(1− βl)p
3l (cost for collision)
0 0.1 0.2 0.3 0.4 0.50
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5x 10
−3
pl
U(p
)
pl* = 0.333
Example: Dependence of a utility function on its own persistence
probability, for βl = 0.5, pmaxl = 0.5, and
Qn∈LI
to(l)(1− pn) = 0.5
![Page 93: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/93.jpg)
Existence of NE
Theorem: There exists NE p∗ characterized by:
p∗l =pmax
l
Qn∈LI
to(l)(1− p∗n)
1− βl(1−Q
n∈LIto(l)(1− p∗n))
An example of the immediate corollaries that confirms intuition:
• Let |LIto(l)| → ∞. If p∗l > 0, then only a finite number of links among
links in LIto(l) have a positive persistence probability at NE
![Page 94: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/94.jpg)
Reverse Engineering BEB Protocol
Is it a gradient-based maximization of Ul(p) over pl?
No, that requires explicit message passing among nodes
Theorem: EB maximizes Ul using stochastic subgradient ascent method
(using only local information on success and collision):
pl(t + 1) = max{pminl , pl(t) + vl(t)}
where
vl(t) = pmaxl 1{Tl(t)=1}1{Cl(t)=0}+βlpl(t)1{Tl(t)=1}1{Cl(t)=1}+pl(t)1{Tl(t)=0}−pl
and
E{vl(t)|p(t)} =∂Ul(p)
∂pl|p=p(t)
![Page 95: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/95.jpg)
Uniqueness and Stability of NE
Example: two links, pmaxl = 1, there are infinite number of NE
Question: under what conditions do we have uniqueness and stability
(convergence of best response strategy) of NE?
Best response strategy:
p∗l (t + 1) = argmaxpmin
l≤pl≤pmax
l
Ul(pl,p∗−l(t))
Theorem: If p∗(0) = pmin,
p∗(2t + 1) → p̂ and p∗(2t) → p̃ as t →∞.
If p̂ = p̃, p̂ is a NE
Proof: S-modular game theory
![Page 96: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/96.jpg)
Uniqueness and Stability of NE
Assume all links have same pmax < 1 and pmin = 0
Let K = maxl{|LIto(l)|}
Uniqueness and stability of NE depend on
• K: amount of potential contention (given)
• β: speed of backoff (variable)
• pmax: minimum amount of backoff (variable)
Theorem: pmaxK4β(1−pmax)
< 1 implies uniqueness and global stability of NE
Proof: Contraction mapping verified through bounding infinity matrix
norm of Jacobian
![Page 97: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/97.jpg)
From Analysis To Design
Motivates forward-engineering:
• How to introduce limited, local message passing of pricing information
to maximize social welfare through selfish utility maximization?
• Can choose utility functions, then design distributed algorithms.
• Prove convergence to global optimum?
Related work:
• Nandagopal et. al. 2000: Conflict graph
• Chen, Low, Doyle 2004: Deterministic approximation
• Kar, Sarkar, Tassiulas 2004: Proportional fair case
![Page 98: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/98.jpg)
Problem Formulation
Nonconvex and coupled generalized NUM:
maximizeP
l∈L Ul(xl)
subject to xl = clplQ
k∈NIto(l)(1− P k), ∀l
Pl∈Lout(n) pl = P n, ∀n
xminl ≤ xl ≤ xmax
l , ∀l0 ≤ P n ≤ 1, ∀n0 ≤ pl ≤ 1, ∀l
variables {xl}, {pl}, {P n}
![Page 99: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/99.jpg)
Algorithm Development
Three main steps:
• Reveal hidden decomposability: log change of variable
• Condition for convexity: application needs to be sufficiently elastic:
Utility function’s curvature needs to be not just negative but bounded
away from 0 by as much as − dUxl (xl)
xldxl
• Standard dual decomposition and distributed subgradient method
![Page 100: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/100.jpg)
Utility-Optimal Random Access
Algorithm 1 (message passing from transmitters):
1: Each node n constructs its local interference graph to obtain sets
Lout(n), Lin(n), LIfrom(n), and NI
to(l), ∀l ∈ Lout(n).
2: Each node n sets t = 0, λl(1) = 1, ∀l ∈ Lout(n),
P n(1) =|Lout(n)|
|Lout(n)|+|LIfrom
(n)| , and pl(1) = 1|Lout(n)|+|LI
from(n)| , ∀l ∈ Lout(n).
3: Locally at each node n, repeat:
3.1: Set t ← t + 1.
3.2: Inform contention prices λl(t) to nodes in NIto(l), ∀l ∈ Lout(n), and
contention probability P n(t) to tl, ∀l ∈ LIfrom(n).
3.3: Set kn(t) =P
l∈Lout(n) λl(t) +P
k∈LIfrom
(n) λk(t) and α(t) = 1t.
![Page 101: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/101.jpg)
3.4: Compute the following:
P n(t + 1) =
8><>:
Pl∈Lout(n) λl(t)P
l∈Lout(n) λl(t)+P
k∈LIfrom
(n)λk(t)
, if kn(t) 6= 0
|Lout(n)||Lout(n)|+|LI
from(n)| , if kn(t) = 0
,
pl(t + 1) =
8><>:
λl(t)Pl∈Lout(n) λl(t)+
Pk∈LI
from(n)
λk(t), if kn(t) 6= 0
1|Lout(n)|+|LI
from(n)| , if kn(t) = 0
,
x′l(t + 1) = argmaxx′minl
≤x′≤x′maxl
˘U ′l (x
′l)− λl(t)x
′l
¯,
and
λl(t + 1) =
264λl(t)− α(t)
0B@c′l + log pl(t) +
X
k∈NIto(l)
log (1− P k(t))− x′l(t)
1CA
375 .
3.5: Each node n decides if it will transmit data with a probability
P n(t). If it decides to transmit data, it chooses to transmit on one of
its outgoing links with probability pl(t)/P n(t), ∀l ∈ Lout(n).
![Page 102: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/102.jpg)
Properties
• Theorem: convergence to global optimum
• Theory accurate predicts performance (unlike deterministic approx.)
• Efficiency-fairness flexible tradeoff
![Page 103: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/103.jpg)
Extensions
• Variant: receiver based message passing (Algorithm 2)
From two-hop message passing to one-hop
• Quantification of message passing overhead
• Message passing reduction heuristics
No need to pass P n
Piggyback λl values to messages in Algorithm 2
Leverage broadcast property in wireless
• Robustness to time delays and outdated messages
![Page 104: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/104.jpg)
Example
1 1.2 1.4 1.6 1.8 21.5
2
2.5
3
3.5
4
α
Net
wor
k ut
ility
Algorithm 1
Algorithm 2
Deterministic approx.
BEB(10,20)
5 5.5 6 6.5 7 7.5 80.8
0.82
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
Rate
Fai
rnes
s
Algorithm 1Algorithm 2Deterministic approx.BEB(10,20)
α = 1
α = 2
![Page 105: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/105.jpg)
Related Results
Reverse engineering:
• Convergence of stochastic subgradient
• Relationship between stochastic subgradient and best response
dynamics
Forward engineering:
• Jointly optimal congestion and contention control
![Page 106: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/106.jpg)
Part IV
Decomposition Theory and Alternative Decompositions
![Page 107: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/107.jpg)
Some References
• D. Palomar and M. Chiang, “A tutorial to decomposition methods for
network utility maximization”, IEEE Journal of Selected Areas in
Communications, vol. 24, no. 8, August 2006
• D. Palomar and M. Chiang, “Alternative decompositions for
distributed maximization of network utility: Framework and
applications”, Proc. IEEE INFOCOM, April 2006
![Page 108: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/108.jpg)
Decomposition
Master Problem
Subproblem 1 Subproblem N...
Original ProblemDecomposition
...
prices/resources
Decompose a problem into subproblems coordinated by a master
problem
Key idea in modularization and distributed control
Mathematical machineries available, but far from complete
![Page 109: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/109.jpg)
Decomposition Theory
• Convexity: efficient solution to global optimality
• Decomposability: distributed algorithm
The two concepts are different
Related in the sense that convexity often leads to zero duality gap, thus
allowing dual decomposition
• No universally-agreed and concise definition of Decomposability
• Can be somewhat quantified by the amount of explicit and implicit
message passing needed (and the growth order as the number of links,
nodes, and processes increase)
![Page 110: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/110.jpg)
Motivation
Daniel Palomar and Mung Chiang 2
![Page 111: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/111.jpg)
Motivation
Daniel Palomar and Mung Chiang 3
![Page 112: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/112.jpg)
Motivation
Daniel Palomar and Mung Chiang 4
![Page 113: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/113.jpg)
Motivation
Daniel Palomar and Mung Chiang 5
![Page 114: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/114.jpg)
Motivation
Daniel Palomar and Mung Chiang 6
![Page 115: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/115.jpg)
Motivation
Daniel Palomar and Mung Chiang 7
![Page 116: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/116.jpg)
Building Blocks
• Primal/Dual decompositions
• Indirect decomposition
• Partial decomposition
• Multilevel/recursive decomposition
• Order of updates: sequential or parallel
• Timescale of updates: iterative or one-shot
• Timescale separation for multilevel decomposition
A variety of combinations from the above building blocks
Example: standard dual algorithm for BNUM is a direct, single-level,
full, dual decomposition
![Page 117: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/117.jpg)
Choices of Decomposition
But there can be many choices of alternative decompositions and
alternative distributed algorithms for GNUM
Standard dual decomposition is not always the best
Even a different representation of GNUM can lead to a different set of
choices of distributed algorithms
Three stages of development:
1. Layering can be understood as decomposition
2. Search through alternative decompositions
3. General methods to exhaust and compare the possibilities
Stages 1 and 2 are done, but not Stage 3
![Page 118: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/118.jpg)
Comparing Decomposition
Metrics (sometimes competing):
• Amount and symmetry of message passing
• Amount and symmetry of local computation
• Speed of convergence (if iterative)
• Robustness (under signaling error, stochastic perturbance, failure of
nodes)
• Ease and robustness of parameter tuning
Some are hard to be quantified/characterized or a full ordering
relationship
Some tradeoffs require application-specific pick among Pareto-optimal
choices of alternative decomposition
![Page 119: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/119.jpg)
Decomposition Techniques: Dual Decomp.
• The dual of the following convex problem (with coupling constraint)
maximize{xi}
∑i fi (xi)
subject to xi ∈ Xi∑i hi (xi) ≤ c
∀i,
is decomposed into subproblems:
maximizexi
fi (xi)− λThi (xi)
subject to xi ∈ Xi.
and the master problem
minimizeλ≥0
g (λ) =∑
i gi (λ) + λTc
where gi (λ) is the optimal value of the ith subproblem.
Daniel Palomar and Mung Chiang 16
![Page 120: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/120.jpg)
Decomposition Techniques: Primal Decomp.
• The following convex problem (with coupling variable y)
maximizey,{xi}
∑i fi (xi)
subject to xi ∈ Xi
Aixi ≤ yy ∈ Y
∀i
is decomposed into the subproblems:maximize
xi∈Xi
fi (xi)
subject to Aixi ≤ y
and the master problemmaximize
y∈Y∑
i f?i (y)
where f?i (y) is the optimal value of the ith subproblem.
Daniel Palomar and Mung Chiang 17
![Page 121: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/121.jpg)
Indirect Primal/Dual Decompositions
• Different problem structures are more suited for primal or dual
decomposition.
• We can change the structure and use either a primal or dual
decomposition for the same problem.
• Key ingredient: introduction of auxiliary variables.
• This will lead to different algorithms for same problem.
Daniel Palomar and Mung Chiang 18
![Page 122: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/122.jpg)
Multilevel Primal/Dual Decompositions
• Hierarchical and recursive application of primal/dual
decompositions to obtain smaller and smaller subproblems:
Master Problem
Subproblem 1...
Secondary Master Problem
prices / resources
First LevelDecomposition
Second LevelDecomposition
Subproblem
Subproblem N
prices / resources
Daniel Palomar and Mung Chiang 19
![Page 123: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/123.jpg)
Applic. 2: Cellular Downlink Power-Rate Control (I)
• Problem:maximize{ri,pi}
∑i Ui (ri)
subject to ri ≤ log (gipi)pi ≥ 0∑
i pi ≤ PT .
∀i
• Decompositions: i) primal, ii) partial dual, iii) full dual.
• Many variants of full dual decomposition: the master problem is
minimizeλ≥0,γ≥0
g (λ, γ)
and can e solved as listed next.
Daniel Palomar and Mung Chiang 27
![Page 124: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/124.jpg)
Applic. 2: Cellular DL Power-Rate Control (II)
1. Direct subgradient update of γ (t) and λ (t)
2. Gauss-Seidel method for g (λ, γ): λ → γ → λ → γ → · · ·
3. Similar to 2), but optimizing λ1 → γ → λ2 → γ → · · ·
4. Additional primal decomp.: minimize g (γ) = infλ≥0 g (λ, γ)
5. Similar to 4), but changing the order of minimization
6. Similarly to 5), but with yet another level of decomposition:
minimize g (λ) sequentially (Gauss-Seidel fashion)
7. Similar to 5) and 6), but minimizing g (λ) with in a Jacobi fashion
Daniel Palomar and Mung Chiang 28
![Page 125: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/125.jpg)
Applic. 2: Cellular DL Power-Rate Control (III)
• Downlink power/rate control problem with 6 nodes with utilities
with utilities Ui (ri) = βi log ri. Evolution of λ4 for all 7 methods:
5 10 15 20 25 30 35 40 45 500
2
4
6
8
10
12
14
16
18
Evolution of λ4 for all methods
iteration
Method 1 (subgradient)Method 2 (Gauss−Seidel for all lambdas and gamma)Method 3 (Gauss−Seidel for each lambda and gamma sequentially)Method 4 (subgradient for gamma and exact for all inner lambdas)Method 5 (subgradient for all lambdas and exact for inner gamma)Method 6 (Gauss−Seidel for all lambdas and exact for inner gamma)Method 7 (Jacobi for all lambdas and exact for inner gamma)
Daniel Palomar and Mung Chiang 29
![Page 126: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/126.jpg)
Part V
Case 1: Joint Congestion Control and Coding
![Page 127: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/127.jpg)
Some References
• J. W. Lee, M. Chiang, and A. R. Calderbank, “Price-based distributed
algorithm for optimal rate-reliability tradeoff in network utility
maximization”, IEEE Journal of Selected Areas in Communications, vol.
24, no. 5, pp. 962-976, May 2006
• M. Chiang, “Balancing transport and physical layer in wireless
multihop networks: Jointly optimal congestion control and power
control,” IEEE Journal of Selected Areas in Communications, vol. 23,
no. 1, pp. 104-116, January 2005
![Page 128: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/128.jpg)
Signal Reliability
Application needs at sources : utility of signal reliability
Physical layer possibilities on links: adaptive coding/modulation
Intuition of the new opportunity:
• Link tradeoff: Fatter pipe, lower reliability
• Source tradeoff: Higher rate, lower quality
Signal quality and physical layer entirely missing from basic NUM
![Page 129: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/129.jpg)
Problem Formulation
Assumptions: decode and reencode with small error probabilities
Reliability: Rs for source s
Code rate: rl,s on link l for source s
Error probability as a function of code rate: El(rl,s)
maximizeP
s Us(xs, Rs)
subject to Rs = 1−Pl∈L(s) El(rl,s), ∀sP
s∈S(l)xsrl,s
≤ Cmaxl , ∀l
xmins ≤ xs ≤ xmax
s , ∀s0 ≤ rl,s ≤ 1, ∀l, s
variables x,R, r
![Page 130: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/130.jpg)
Overview
Difficulty: Neither convex nor separable problem
Goal: Derive globally optimal and distributed algorithm
• Develop such algorithms
• Extend pricing interpretation
• Sufficient conditions for convergence to global optimum
• Techniques to tackle nonconvexity and nonseparability issues
![Page 131: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/131.jpg)
Integrated Policy
Each link maintains the same code rate for all sources traversing it
maximizeP
s Us(xs, Rs)
subject to Rs ≤ 1−Pl∈L(s) El(rl), ∀sP
s∈S(l)xsrl≤ Cmax
l , ∀lvariables x,R, r
Naturally decompose:
X
s∈S(l)
xs ≤ Cmaxl rl, ∀l
![Page 132: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/132.jpg)
Difficulty 1: Nonconvexity
Approximation of El(rl):
pl ≤ exp(−NE0(rl))
E0(rl) = max0≤ρ≤1
maxQ
[Eo(ρ,Q)− ρrl]
Eo(ρ,Q) = − log
J−1X
j=0
"K−1X
k=0
Q(k)P (j|k)1/(1+ρ)
#1+ρ
Lemma:
If absolute value of first derivatives of E0(rl): bounded away from 0,
Absolute value of second derivative of E0(rl): upper bounded,
Then for a large enough codeword block length N , El(rl) is a convex
function
![Page 133: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/133.jpg)
Standard Methodology
Next: Use standard Lagrangian relaxation and distributed subgradient
algorithm to develop distributed algorithm
![Page 134: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/134.jpg)
Distributed Algorithm 1
Source problem and reliability price update at source s:
• Source problem:
maximize Us(xs, Rs)− λs(t)xs − µs(t)Rs
subject to xmins ≤ xs ≤ xmax
s
where λs(t) =P
l∈L(s) λl(t) is the end-to-end congestion price at
iteration t
• Reliability price update:
µs(t + 1) = [µs(t)− α(t) (Rs(t)−Rs(t))]+
where Rs(t) = 1−Pl∈L(s) El(rl(t)) is the end-to-end reliability at
iteration t
![Page 135: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/135.jpg)
Link problem and congestion price update at link l:
• Link problem:
maximize λl(t)rlCmaxl − µl(t)El(rl)
subject to 0 ≤ rl ≤ 1
where µl(t) =P
s∈S(l) µs(t) is the aggregate reliability price paid by
sources using link l at iteration t
• Congestion price update:
λl(t + 1) =hλl(t)− α(t)
“rl(t)C
maxl − xl(t)
”i+
where xl(t) =P
s∈S(l) xs(t) is the aggregate information rate on link l at
iteration t
![Page 136: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/136.jpg)
Pricing Interpretation
• Source problem: maximize total net utility on
rate (with total congestion price) and
reliability (with signal quality price)
• Source algorithm:
local solution of source problem (2 variables)
updates signal quality price
• Network problem: maximize net revenue:
receive revenue from rate
pay price for unreliability
• Link algorithm: update link congestion price
![Page 137: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/137.jpg)
Distributed Algorithm 1
Source s Link l
sx
sµ
lλ∑∈
=)(sLl
ls λλ
lr∑∈
−=)(
)(1sLl
lls rER
∑∈
=)(lSs
sl µµ
∑∈
=)(lSs
sl xx
Network
Network
sR
Theorem: Distributed Algorithm 1 converges to the globally optimal
rate-reliability tradeoff for sufficiently strong codes
![Page 138: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/138.jpg)
Differentiated Policy
Each link may give a different code rate for each of the sources
traversing it
• Per-flow state needed
• Better rate-reliability tradeoff
maximizeP
s Us(xs, Rs)
subject to Rs ≤ 1−Pl∈L(s) El(rl,s), ∀sP
s∈S(l)xsrl,s
≤ Cmaxl , ∀l
variables x,R, r
![Page 139: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/139.jpg)
Difficulty 2: Coupling
Step 1: Introduce auxiliary variables (a new scheduling layer):
maximizeP
s Us(xs, Rs)
subject to Rs ≤ 1−Pl∈L(s) El(rl,s), ∀sxsrl,s
≤ cl,s, ∀l, s ∈ S(l)P
s∈S(l) cl,s ≤ Cmaxl , ∀l
Step 2: Log change of variables (on x):
maximizeP
s U ′s(x′s, Rs)
subject to Rs ≤ 1−Pl∈L(s) El(rl,s), ∀sx′s − log rl,s ≤ log cl,s, ∀l, s ∈ S(l)P
s∈S(l) cl,s ≤ Cmaxl , ∀l
Separable problem but U ′s(x′s, Rs) may not be concave
![Page 140: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/140.jpg)
Difficulty 2: Coupling
Step 3: Concavity condition
gs(xs, Rs) =∂2Us(xs, Rs)
∂x2s
xs +∂Us(xs)
∂xs,
hs(xs, Rs) =
„∂2Us(xs, Rs)
∂xs∂Rs
«2
−∂2Us(xs, Rs)
∂x2s
∂2Us(xs, Rs)
∂R2s
«xs
−∂2Us(xs, Rs)
∂R2s
∂Us(xs, Rs)
∂xs,
and
qs(xs, Rs) =∂2Us(xs, Rs)
∂R2s
.
Lemma: If gs(xs, Rs) < 0, hs(xs, Rs) < 0, and qs(xs, Rs) < 0, then
U ′s(x′s, Rs) is a concave function of x′s and Rs
![Page 141: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/141.jpg)
Difficulty 2: Coupling
Special case 1: α-fair utilities
Us(xs, Rs) =
8<:
log xsRs, if α = 1
(1− α)−1(xsRs)1−α, otherwise
If α ≥ 1, conditions for concavity is satisfied
Special case 2: if Us is additive in xs and Rs, its curvature needs to be
not just negative but bounded away from 0 by as much as − dUxs (xs)
xsdxs
The application traffic is sufficiently elastic
![Page 142: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/142.jpg)
Distributed Algorithm 2
All descriptions same as in Algorithm 1 except one:
• Link problems:
Bandwidth allocation problem
maximizeP
s∈S(l) λl,s(t) log cl,s
subject toP
s∈S(l) cl,s ≤ Cmaxl
Code rate allocation problem for source s, s ∈ S(l)
maximize λl,s(t) log rl,s − µs(t)El(rl,s)
subject to 0 ≤ rl,s ≤ 1
![Page 143: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/143.jpg)
Numerical Example
U1 and U2U5 and U6
U3 and U4 U7 and U8
L1 L2 L3 L4
Us(xs, Rs) = asx1−α
s − xmin(1−α)s
xmax(1−α)s − x
min(1−α)s
+ (1− as)R
(1−α)s −R
min(1−α)s
Rmax(1−α)s −R
min(1−α)s
Example 1: as = a
Example 2: as =
8<:
0.5− v, if s is an odd number
0.5 + v, if s is an even number,
![Page 144: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/144.jpg)
Numerical Example 1
0.1 0.2 0.3 0.4 0.50.9
0.92
0.94
0.96
0.98
1
Data rate (Mbps)
Rel
iabi
lity
User 1User 3User 5
a=0
a=1
0 0.2 0.4 0.6 0.8 12
3
4
5
6
7
8
aN
etw
ork
utili
ty
Diff. reliabilityInt. reliabilityStat. reliability
![Page 145: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/145.jpg)
Numerical Example 2
0.1 0.15 0.2 0.25 0.3 0.350.9
0.92
0.94
0.96
0.98
1
Data rate (Mbps)
Rel
iabi
lity
User 2 (diff.)User 2 (int.)User 2 (stat.)
v=0
v=0.5
0 0.1 0.2 0.3 0.4 0.52.5
3
3.5
4
4.5
5
5.5
6
vN
etw
ork
utili
ty
Diff. reliability
Int. reliability
Stat. reliability
![Page 146: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/146.jpg)
Extensions
Partial Dynamic Reliability Policy
• Only some links can adjust error correction capability
• DSL links in end-to-end path
• Substantial gain still observed.
Wireless MIMO rate reliability control
• Multiplexing gain vs. diversity gain
• Ignore multi-point joint coding
• Sufficiently high SNR (for convexity)
• Same mathematical structures as before
![Page 147: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/147.jpg)
Key Messages and Methods
• Convexity of the generalized NUM is the key to devising a globally
optimal solution.
Conditions on constraints and utility curvature for convexity to hold
• Decomposability of the generalized NUM is the key to devising a
distributed solution.
Introducing new “layer” and log change of variable to reveal hidden
decomposability structure
• User-generated pricing following end-to-end principle
![Page 148: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/148.jpg)
Part VI
Case 2: Joint Congestion Control, Routing, and Scheduling
![Page 149: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/149.jpg)
Some References
• L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Joint optimal
congestion control, routing, and scheduling in wireless ad hoc
networks,” Proc. IEEE INFOCOM, April 2006
• A. Eryilmaz and R. Srikant, “Fair resource allocation in wireless
networks using queue-length-based scheduling and congestion control,”
Proc. IEEE INFOCOM, March 2005
• X. Lin and N. Shroff, “The impact of imperfect scheduling on
cross-layer rate control in wireless networks,” Proc. IEEE INFOCOM,
March 2005
• M. Neely, E. Modiano, and C. Rohrs, “Dynamic power control and
routing over time varying wireless networks”, IEEE Journal of Selected
Areas in Communications, vol. 23, no. 1, pp. 89-103, January 2005
• A. L. Stolyar, “Maximizing queueing network utility subject to
statbility: greedy primal-dual algorithm”, Queueing Systems, vol. 50,
no. 4, pp. 401-457, 2005
![Page 150: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/150.jpg)
Network Model
Wireless network: a set of nodes and a set of logical linksEach link has fixed capacity when active
Primary interference: links that share common node cannot transmit simultaneously
NL
Ll∈ lc
![Page 151: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/151.jpg)
Schedulability
Independent set: links that can transmit simultaneously
An independent set e is represented by an L-dim rate vector re:
Feasible rate region is:
⎩⎨⎧ ∈
=otherwise
if0
elcr le
l
⎭⎬⎫
⎩⎨⎧
=≥==Π ∑∑ 1,0,:e
ee
e
ee aararr
![Page 152: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/152.jpg)
i j
dijf
Schedulability constraint
: capacity of link (i, j) allocated to flow with destination d
: capacity of link (i, j) allocated to all flows traversing link (i, j)
Schedulability constraint:
dijf
( )),( links ,:
:
jiff
ff
ij
d
dijij
∀=
=∑
Π∈ f
'dijf
ijf
![Page 153: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/153.jpg)
DdNiffxj
dij
j
dji
di ∈∈≤+ ∑∑ , allfor
Rate constraint
i∑∈Ljij
dijf
),(:∑
∈Ljij
djif
),(:
dix
capacity allocated to incoming transit flows
capacity allocated to all outgoing flows
external flow entering at node i
![Page 154: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/154.jpg)
Problem formulation
Generalized NUM:
Π∈
≠−≤ ∑∑
∑
f
diffxts
xU
j
dji
j
dij
di
di
diid
fx dijid
, ..
)( max,,
![Page 155: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/155.jpg)
Dual decomposition
Dual decomposition:
Subgradient
Subgradient algorithm to min D(p)
Π∈
+−−= ∑ ∑∑∑fts
ffxpxUpDdi j
dji
j
dij
di
di
di
diid
fx dijid
..
)()(max)(,,,
∑∑ −+=j
dij
j
dji
di
di pfpfpxpg )()()()(
[ ]++=+ ))(()()1( tpgtptp di
di
di γ
![Page 156: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/156.jpg)
Dual decomposition
Dual decomposition:
Dual problem has 2 subproblems:
Π∈
−= ∑∑ ∑f
ffppDj
dji
di j
dij
di
f dij
s.t.
)(max)(,
2
di
di
di
diid
xpxxUpD
di
−= ∑,
1 )( max)( rate control
routing,scheduling
Π∈
+−−= ∑ ∑∑∑fts
ffxpxUpDdi j
dji
j
dij
di
di
di
diid
fx dijid
..
)()(max)(,,,
![Page 157: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/157.jpg)
Cross-layer implementation
1st subproblem: solved by rate control using local congestion price
))(()( 1' tpUtx diid
di
−=
Transport: rate control based on local price
( )∑ −=di
di
di
diid
xpxxUpD
di,
1 )(max )(
![Page 158: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/158.jpg)
Cross-layer implementation
2nd subproblem: equivalent form solved by routing and scheduling
Nodes maintain a separate queue for each destination d
)()(maxarg:)(
)()(max :)(
tptptd
tptptwdj
didij
dj
didij
−=
−=max diff. price:
dest with max wij(t):
( )Π∈
−= ∑f
ppfpD dj
didji
ijf ij
s.t.
maxmax)(,
2
routing
scheduling
![Page 159: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/159.jpg)
Cross-layer implementation
2nd subproblem: equivalent form solved by routing and scheduling
Network: routing based on differential priceOutput link (i, j) serves only queue dij(t) with max differential priceIt transmits from queue dij(t) at rate cl if it is scheduled to send
( )Π∈
−= ∑f
ppfpD dj
didji
ijf ij
s.t.
maxmax)(,
2
routing
scheduling
![Page 160: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/160.jpg)
Cross-layer implementation
2nd subproblem: equivalent form solved by routing and scheduling
Link: scheduling (centralized)Solve for maximum independent set e(t):
All links l in e(t) transmit at rates cl
)(maxarg:)(),(∑
∈∈=
ejiijij
Eetwcte
( )Π∈
−= ∑f
ppfpD dj
didji
ijf ij
s.t.
maxmax)(,
2
routing
scheduling
Becomes stochastic with time-varying channels
![Page 161: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/161.jpg)
Summary of Cross-layer Algorithm
)(txdi
Queue at nodes
TCP rate control
Routing +Scheduling
)( tf dij
[ ]++=+ ))(()()1( tpgtptp di
di
di γ
)(tpdi )(tpd
i
![Page 162: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/162.jpg)
Stability
TheoremThe subgradient algorithm converges
arbitrarily close to optimum
i.e., given any δ>0, there exists a sufficiently small stepsize such that
δ
δ
+≤
−≥
∞→
∞→
)())(( suplim
)())(( inflim*
*
pDtpD
xPtxP
t
t
primal objective function
dual objective function
optimum : ,avg running :)( ),(
** pxtptx
![Page 163: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/163.jpg)
Time-varying Channel
Channel state is an i.i.d. finite state process with distribution In channel state
Link l’s capacity is cl(h)Feasible rate region is
Extend the cross-layer algorithm with only a modification to scheduling
Questions: stability? optimality?
h))(( thq
)(th
)(hΠ
))(( s.t. )(max,
thftwf ijji
ijf ij
Π∈∑ random
![Page 164: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/164.jpg)
Reference SystemDefine mean feasible rate region
Define reference system problem
⎭⎬⎫
⎩⎨⎧
Π∈==Π ∑ )()( ,)()(: hhrhrhqrrh
Π∈
≠−≤ ∑∑
∑
f
diffxts
xU
j
dji
j
dij
di
di
diid
fx dij
di
, ..
)( max,,
![Page 165: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/165.jpg)
StabilityCongestion price is a positively recurrent Markov chain
Proof by stochastic Lyapunov analysisDual function of the reference problem:Optimal price: Lyapunov function:Then:
where
)(pD*p
22||||)( ∗−= pppV
( )ApApIIGptptpVtpVE c ∈∈
−−≤=−+ 22])(|))(())1(([ γ
}||:||{ 2 δ≤−= ∗pppA
[ ]++=+ ))(()()1( tpgtptp dit
di
di γ
2)()(
||||max2
∗
≤−−=
∗pp
GpDpD γδ
![Page 166: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/166.jpg)
Optimality
TheoremThe stochastic subgradient algorithm
converges arbitrarily close to optimum
i.e., given any δ>0, there exists a sufficiently small stepsize such that
[ ] [ ]problem reference of optimum : ,
statesteady in vaueexpected :)( ,)(** px
pExE ∞∞
( )( ) δ
δ
−≥∞
+≤∞
)()]([)()]([
*
*
xPxEPpDpED
![Page 167: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/167.jpg)
Conclusion
Joint rate control, routing, scheduling design for wireless networksSubgradient algorithm
Node prices adjusted according to excess demandTraffic source controls its rate using marginal utility function based on local priceOnly destination (queue) with maximum differential price is served over each linkRouting ‘absorbed’ into congestion control and scheduling Combine backpressure and congestion pricing
Extension to time-varying channelIn general: dual solution to convex G.NUM remains stable and optimal (on average) under (Markov model) stochastically varying constraint set
![Page 168: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/168.jpg)
Part VII
Case 3: TCP/IP Interactions
![Page 169: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/169.jpg)
Some References
• J. Wang, L. Li, S. H. Low and J. C. Doyle, “Cross-layer Optimization
in TCP/IP Networks,” IEEE/ACM Transactions on Networking, vol.
13, no. 3, pp. 582-268, June 2005
• J. He, M. Chiang, and J. Rexford, “TCP/IP interaction based on
congestion prices: Stability and optimality”, Proc. IEEE ICC, June 2006
![Page 170: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/170.jpg)
Protocol decomposition
TCP-AQM:TCP algorithms maximize utility with different utility functions
Congestion prices coordinate across protocol layers
BccRx
xU
Ti
iixRc
≤≤
∑≥
α , subject to
)( maxmax max0
TCP-AQM
IP
TCP/AQM
Applications
Link
![Page 171: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/171.jpg)
Protocol decomposition
TCP-AQMIP
BccRx
xU
Ti
iixRc
≤≤
∑≥
α , subject to
)( maxmax max0
TCP-AQM
IP
TCP/AQM
Applications
Link
TCP/IP:TCP algorithms maximize utility with different utility functionsShortest-path routing is optimal using congestion prices as link costs ……
Congestion prices coordinate across protocol layers
![Page 172: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/172.jpg)
Two timescales
Instant convergence of TCP/IPLink cost = a pl(t) + b dl
Shortest path routing R(t)price
static
TCP/AQM
IPR(0)
a p(0)
R(1)
a p(1)
… R(t), R(t+1) ,…
![Page 173: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/173.jpg)
TCP-AQM/IP Model
TCP
cxtR
xUtxi
iix
≤
= ∑≥
)( subject to
)( max arg )(0
∑
∑ ∑
+
⎟⎠
⎞⎜⎝
⎛−=
≥≥
lll
i llliiiixp
pc
ptRxxUtpi
)()(max min arg )(00
AQM
))((minarg)1( lll
liRi bdtapRtRli
+=+ ∑IP
Link cost
![Page 174: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/174.jpg)
Questions
Does equilibrium routing Ra exist ?What is utility at Ra?Is Ra stable ? Can it be stabilized?
TCP/AQM
IPR(0)
a p(0)
R(1)
a p(1)
… R(t), R(t+1) ,…
![Page 175: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/175.jpg)
Delay-insensitive utility Ui(xi)
TheoremTCP/IP equilibrium solves the primal problem
for general networks only if b=0i.e., if route based purely on congestion prices
⎟⎟⎠
⎞⎜⎜⎝
⎛+⎟
⎠
⎞⎜⎝
⎛−
≤
∑∑ ∑
∑
≥≥
≥
ll
li l
lliRiiixp
iii
xR
cppRxxU
cRxxU
ii
max)( max min
subject to )( maxmax
00
0
:Dual
:Primal
![Page 176: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/176.jpg)
TheoremIf b=0, Ra exists iff zero duality gap
Shortest-path routing is optimal with congestion pricesNo penalty for not splitting
⎟⎟⎠
⎞⎜⎜⎝
⎛+⎟
⎠
⎞⎜⎝
⎛−
≤
∑∑ ∑
∑
≥≥
≥
ll
li l
lliRiiixp
iii
xR
cppRxxU
cRxxU
ii
max)( max min
subject to )( maxmax
00
0
:Dual
:Primal
Delay-insensitive utility Ui(xi)
![Page 177: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/177.jpg)
TCP-AQMIP
TCP/IP: b=0Equilibrium of TCP/IP exists iff zero duality gapNP-hard, but subclass with zero duality gap is in P Equilibrium, if exists, can be unstableCan stabilize, but with reduced utility
BccRx
xU
Ti
iixRc
≤≤
∑≥
α , subject to
)( maxmax max0
TCP-AQM
IP
TCP/AQM
Applications
Link
Delay-insensitive utility Ui(xi)
![Page 178: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/178.jpg)
Delay-sensitive utility Ui(xi, di)
TheoremIf a>0, b>0, then TCP/IP equilibrium solves the
primal problem for general networks if
Moreover no other “reasonable” class of utility functions work
⎟⎟⎠
⎞⎜⎜⎝
⎛+⎟⎟
⎠
⎞⎜⎜⎝
⎛−⎟⎠
⎞⎜⎝
⎛
≤⎟⎠
⎞⎜⎝
⎛
∑∑ ∑∑
∑ ∑
≥≥
≥
ll
li l
lliRil
lliiixp
i llliii
xR
cppRxRxU
cRxRxU
ii
max, max min
subject to , maxmax
00
0
:Dual
:Primal
τ
τ
iiiiiii dxabxVdxU −= )( ),(
![Page 179: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/179.jpg)
TheoremSuppose a>0, b>0 and Then equilibrium routing exists iff zero duality gap
Shortest-path routing is optimal with congestion pricesNo penalty for not splitting
iiiiiii dxbaxVdxU 1)( ),( −−=
Delay-sensitive utility Ui(xi, di)
⎟⎟⎠
⎞⎜⎜⎝
⎛+⎟⎟
⎠
⎞⎜⎜⎝
⎛−⎟⎠
⎞⎜⎝
⎛
≤⎟⎠
⎞⎜⎝
⎛
∑∑ ∑∑
∑ ∑
≥≥
≥
ll
li l
lliRil
lliiixp
i llliii
xR
cppRxRxU
cRxRxU
ii
max, max min
subject to , maxmax
00
0
:Dual
:Primal
τ
τ
![Page 180: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/180.jpg)
Extensions
Other timescale separations between TCP and IP dynamics (He, Chiang, Rexford 2006)Forward engineering:
DATE (Dynamic Adaptive Traffic Engineering)
HTTP/TCP interactions (Chang Liu 2004)
![Page 181: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/181.jpg)
Part VIII
Future Research Challenges and Summary
![Page 182: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/182.jpg)
Future Research Issues
• Technical: Global stability under delay...
• Modeling: routing in ad hoc network, ARQ, MIMO...
• Time issues
• Why deterministic fluid model?
Shannon 1948: remove finite blocklength, Law of Large Numbers kicks
in (later finite codewords come back...)
Kelly 1998: remove coupled queuing dynamics, optimization and
decomposition view kicks in (later stochastics come back...)
• What if it’s not convex optimization?
Rockafellar 1993: Convexity is the watershed between easy and hard
(what if it’s hard?)
• Is performance the only optimization objective?
![Page 183: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/183.jpg)
Research Challenges
A sample of 30 bullets in three categories
![Page 184: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/184.jpg)
Open Problems
• Stochastic stability for general filesize distribution, general utility
functions and convex constraint set, without timescale separation?
• Performance (utility, delay...) under session, channel, and packet level
stochastic?
• Impacts of stochastic feedback for multi-timescale decompositions?
• Validation of fluid model from packet level dynamics?
• Global convergence of successive convex approximations for signomial
programming?
• Distributed Sum-of-Squares for nonconcave NUM
• Duality gap: estimation, bounding, and implications
• Tight bound on the rate of convergence of various distributed
algorithms?
• Practical stepsize rules in asynchronous networks?
• Low spatial-temporal complexity scheduling algorithm?
• Global stability under feedback delay?
![Page 185: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/185.jpg)
Open Issues
• Constraint set of G.NUM from information theory?
• How to systematically search alternative G.NUM representations and
alternative decompositions?
• Adaptive slicing by primal decompositions?
• Modeling of routing (ad hoc network and BGP)?
• Dealing with utility as functions of delay and transient resource
allocations for real-time flows?
• Degree of heterogeneity and price of heterogeneity?
• Topology level stochastic?
• New notions of fairness in S.NUM?
• Quantify suboptimality’s impact on fairness?
• Characterize and bound instability?
• Hardware and application modeling?
![Page 186: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/186.jpg)
New Mentalities
• Robustness-optimality tradeoff?
• Move away from optimality?
Suboptimal (with bounded loss of optimality) and simple algorithm for
each module
Good architecture contains the “damage” to the overall system
• Stochastic network dynamics is good?
“Washes away” the corner cases?
• From focus on equilibrium to investigations of the transients (eg, how
close to optimum within a given time, will resource allocation during
transient drop below certain thresholds)?
• How to compare alternative architectures?
• Redesign architectures (especially the division between control
protocols and network management systems) for optimizability?
• Quantify other Network X-ities?
• Managing complexity in networks through layering?
![Page 187: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/187.jpg)
A Sample of 20 Methodologies
• Reverse engineering cooperative protocol as an optimization algorithm
• Lyapunov function construction to show stability
• Proving convergence of dual descent algorithm
• Proving stability by singular perturbation theory
• Proving stability by passivity argument
• Proving equilibrium properties through vector field representation
• Reverse engineering non-cooperative protocol as a game
• Verifying contraction mapping by bounding the Jacobian’s norm
• Analyzing cross-layer interaction systematically through G.NUM
• Change of variable for decoupling, and computing minimum curvature
needed
![Page 188: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/188.jpg)
A Sample of 20 Methodologies
• Dual decomposition for jointly optimal cross layer design
• Computing conditions under which a general constraint set is convex
• Introducing an extra “layer” to decouple the problem
• End user generated pricing
• Different timescales of protocol stack interactions through different
decomposition methods
• Maximum differential congestion pricing for node-based back-pressure
scheduling
• Absorbing routing functionality into congestion control and scheduling
• Primal and dual decomposition for coupling constraints
• Consistency pricing for decoupling coupled objective
• Partial and hierarchical decompositions for architectural alternatives
![Page 189: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/189.jpg)
10 Key Messages
• Existing protocols in layers 2,3,4 have been reverse engineered
• Reverse engineering leads to better design
• There is one unifying approach to cross-layer design
• Loose coupling through layering price
• Queue length often a right layering price, but not always
• Many alternatives in decompositions and layering architectures
• Convexity is key to proving global optimality
• Decomposability is key to designing distributed solution
• Still many open issues in modeling, stochastic dynamics, and
nonconvex formulations
• Architecture, rather than optimality, is the key
![Page 190: Layering As Optimization Decomposition: A Mathematical Theory of](https://reader031.vdocuments.net/reader031/viewer/2022022416/586cfa631a28ab2a738c5cb2/html5/thumbnails/190.jpg)
Contacts
www.princeton.edu/∼chiangm
netlab.caltech.edu