Theory of Complex Networks John Doyle Control and Dynamical Systems Caltech Transportation Finance Health Commerce Our lives are run by/with networks Energy Consumer Emergency Manufacturing Information Utilities.

Download Report

Transcript Theory of Complex Networks John Doyle Control and Dynamical Systems Caltech Transportation Finance Health Commerce Our lives are run by/with networks Energy Consumer Emergency Manufacturing Information Utilities.

Theory of
Complex
Networks
John Doyle
Control and Dynamical Systems
Caltech
Transportation
Finance
Health
Commerce
Our lives are run
by/with networks
Energy
Consumer
Emergency
Manufacturing
Information
Utilities
Environment
Health Transportation Finance
Commerce Convergent
Energy
Emergency
Networks Consumer
Manufacturing
Utilities
Information
Convergent networking: the promise
Ubiquitous computing,
communications, and
control
• that is embedded and
intertwined
• via sensors and actuators
• in complex networks of
networks, with layers of
protocols and feedback.
Resulting in:
• Seamless integration and
automation of everything
• Efficient and economic
operation
• Robust and reliable services
Environment
Health Transportation Finance
Commerce
Energy
Convergent
Emergency
Networks
Consumer
Manufacturing
Information
Utilities
Convergent networking: the reality
• Right now, back in Los Angeles, we can experience (in
addition to smog, earthquakes, fires, floods, riots, lawyers,…)
– Widespread and prolonged power outages from lightning strikes in
Washington (or just “nonequilibrium market fluctuations”).
– Widespread and prolonged flight delays from weather or ATC software
glitches in Chicago or Atlanta.
– Internet meltdowns caused by hackers in Moscow.
– Financial meltdowns caused by brokers in Singapore.
• What can we expect?
– Widespread and prolonged meltdowns of integrated power,
transportation, communication, and financial networks caused by
lightning strikes in Singapore or a new release of MS Windows 2020?
Elements of systems
• Sense the environment
and internal state
• Extract what’s novel
• Communicate or store
what’s novel
• Extract what’s useful
• Compute decisions based
on what’s useful
• Take action
• Evaluate consequences
• Repeat
We want results
H
A
R
D
E
R
Data
Is not novel information
Is not useful Information
Is not knowledge
Is not understanding
Is not wisdom
Is not action
Is not results
Two great abstractions of the 20th Century
1. Separate systems engineering into control,
communications, and computing
–
–
Theory
Applications
2. Separate systems from physical substrate
• Facilitated massive, wildly successful, and explosive
growth in both mathematical theory and
technology…
• …but creating a new Tower of Babel where even the
experts do not read papers or understand systems
outside their subspecialty.
Tower of Babel
• Issues for theory
– Rigor
– Relevance
– Accessibility
• Spectacular success on the first two
• Little success on the last one, which is critical for a
multidisciplinary approach to systems biology
• Perhaps all three is impossible?
• (In contrast, there are whole research programs in
“complex systems” devoted exclusively to
accessibility. They have been relatively “popular,”
but can be safely ignored in biology.)
Biology and advanced technology
• Biology
– Integrates control, communications, computing
– Into distributed control systems
– Built at the molecular level
• Advanced technologies will do the same
• We need new theory and math, plus
unprecedented connection between systems and
devices
• Two challenges for greater integration:
– Unified theory of systems
– Multiscale: from devices to systems
Compute
Communications and computing
Compute
Act
Sense
Environment
Computation
Devices
Devices
Control
Dynamical Systems
From
• Software to/from human
• Human in the loop
Compute
To
• Software to Software
• Full automation
• Integrated control,
comms, computing
• Closer to physical
substrate
Computation
• New capabilities & robustness
• New fragilities & vulnerabilities
Devices
Devices
Control
Dynamical Systems
Theoretical foundations
•
•
•
•
•
Computational complexity: decidability, P-NP-coNP
Information theory: source and channel coding
Control theory: feedback, optimization, games
Dynamical systems: dynamics, bifurcation, chaos
Statistical physics: phase transitions, critical phenomena
• Unifying theme: uncertainty management
• Different abstractions and relaxations
• Integrating these theories involves new math, much not traditionally
be viewed as “applied,” e.g..
– Perturbation theory of operator Banach algebras
– Semi-algebraic geometry
Uncertainty management
• Each domain faces similar abstract issues and
tradeoffs, but with differing details:
• Sources of uncertainty
• Limited resources
• Robust strategies
• Fundamental tradeoffs
• Ignored issues
Control theory
• Sources of uncertainty: plant uncertainty and
sensor noise
• Limited resources: sensing, actuation, and
computation
• Robust strategies: feedback control and related
methods
• Fundamental tradeoffs: Bode’s integral formula,
RHP zeros, saturations, …
• Ignored issues: communications in distributed
control, software reliability
Information theory
• Sources of uncertainty: source and channel
• Limited resources: storage, bandwidth, and
computation
• Robust strategies: coding
• Fundamental tradeoffs: capacity, rate-distortion
• Ignored issues: feedback and dynamics
Computation complexity
• Sources of uncertainty: intractability, problem
instance
• Limited resources: computer time and space
• Robust strategies: algorithms
• Fundamental tradeoffs: P/NP/Pspace/undecidable
• Ignored issues: real-time, uncertainty in physical
systems
Software correctness
•
•
•
•
•
Sources of uncertainty: bugs, user inputs
Limited resources: computer time and space
Robust strategies: formal verification
Fundamental tradeoffs: computational complexity
Ignored issues: real-time, uncertainty in physical
systems
Multiscale physics
• Sources of uncertainty: initial conditions,
unmodeled dynamics, quantum mechanics
• Limited resources: computer time and space,
measurements
• Robust strategies: coarse graining,
renormalization??
• Fundamental tradeoffs: energy/matter, entropy,
quantum, etc…
• Ignored issues: robustness, rigor, computation, etc
• (This looks mostly fixable.)
Unified theory of uncertainty management
• Sources of uncertainty: plant, multiscale physics,
sensors, channels, bugs, user inputs
• Limited resources: computer time and space,
energy, materials, bandwidth, actuation
• Robust strategies: ??
• Fundamental tradeoffs: ??
• Ignored issues: human factors
Progress
• Unified view of web and internet protocols
–
–
–
–
–
Good place to start
Add feedback and dynamics to communications
Observations: fat tails (Willinger)
Theory: Source coding and web layout (Doyle)
Theory: Channel coding and congestion control (Low)
• Unified view of robustness and computation
– Anecdotes from engineering and biology
– New theory (especially Parrilo)
– Not enough time today…
Bonus!
• “Unified systems” theory helps resolve fundamental
unresolved problems at the foundations of physics
• Ubiquity of power laws (statistical mechanics)
• Shear flow turbulence (fluid dynamics)
• Macro dissipation and thermodynamics from micro
reversible dynamics (statistical mechanics)
• Quantum-classical transition
• Quantum measurement
• Thus the new mathematics for a unified theory of
systems is directly relevant to multiscale physics
• The two challenges (unify and multiscale) are
connected.
Network protocols.
HTTP
Files
TCP
IP
packets
packets
packets
packets
packets
packets
Routers
web traffic
Web/internet traffic
Is streamed out
on the net.
Web
servers
Creating
internet traffic
Web
client
web traffic
Is streamed
out on the net.
Web
servers
Creating
internet traffic
Let’s look at
some web traffic
Web
client
6
5
Frequency
(Huffman)
(Crovella)
4
Cumulative
Data
compression
WWW files
Mbytes
3
Forest fires
1000 km2
2
(Malamud)
1
Los Alamos fire
0
-1
-6
-5
Decimated data
Log (base 10)
-4
-3
-2
-1
0
1
Size of events
2
6
Web files
5
Codewords
4
Cumulative
Frequency
-1
3
Fires
2
-1/2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
Size of events
Log (base 10)
2
6
5
Frequency
(Huffman)
(Crovella)
4
Cumulative
Data
compression
WWW files
Mbytes
3
Forest fires
1000 km2
2
(Malamud)
1
Los Alamos fire
0
-1
-6
-5
Decimated data
Log (base 10)
-4
-3
-2
-1
0
1
Size of events
2
20th Century’s 100 largest disasters worldwide
2
10
Technological ($10B)
Natural ($100B)
1
10
US Power outages
(10M of customers)
0
10
-2
10
-1
10
0
10
2
10
Log(Cumulative
frequency)
1
10
= Log(rank)
0
10
-2
10
-1
10
Log(size)
0
10
100
80
Technological ($10B)
rank
60
Natural ($100B)
40
20
0
0
2
4
6
8
size
10
12
14
2
100
10
Log(rank)
1
10
10
3
2
0
1
10
-2
10
-1
0
10
10
Log(size)
20th Century’s 100 largest disasters worldwide
2
10
Technological ($10B)
Natural ($100B)
1
10
US Power outages
(10M of customers)
Slope = -1
(=1)
0
10
-2
10
-1
10
0
10
6
Data
compression
WWW files
Mbytes
5
4
Cumulative
Frequency
-1
3
Forest fires
1000 km2
2
-1/2
1
0
-1
-6
-5
Decimated data
Log (base 10)
-4
-3
-2
-1
0
1
Size of events
2
6
5
4
Cumulative
Frequency
Data
compression
WWW files
Mbytes
exponential
-1
3
Forest fires
1000 km2
2
-1/2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
Size of events
2
0
0
10
1
10
2
3
10
10
10
.5
-1
10
-2
10
loglog
1
semilogy
-3
10
exp
-4
10
1
Plotting
power laws
and exponentials
linear
0.6
0.2
20
40
60
80
100
6
Data
compression
WWW files
Mbytes
5
exponential
4
Cumulative
Frequency
3
Forest fires
1000 km2
2
All events are
close in size.
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
Size of events
2
6
5
4
Cumulative
Frequency
Data
compression
WWW files
Mbytes
-1
3
Forest fires
Most2events
1000 km2
are small
1
0
-1/2
But the large
events are huge
-1
-6
-5
-4
-3
-2
-1
0
1
Size of events
2
6
5
4
Cumulative
Frequency
Data
compression
WWW files
Robust
Mbytes
-1
3
Forest fires
Most2events
1000 km2
are small
1
0
-1/2
Yet
Fragile
But the large
events are huge
-1
-6
-5
-4
-3
-2
-1
0
1
Size of events
2
Robustness of
HOT systems
Fragile
Robust
(to known and
designed-for
uncertainties)
Fragile
(to unknown
or rare
perturbations)
Robust
Uncertainties
Large scale phenomena is extremely
non-Gaussian
• The microscopic world is largely exponential
• The laboratory world is largely Gaussian
because of the central limit theorem
• The large scale phenomena has heavy tails
(fat tails) and power laws
Size of events x vs. frequency
dP
 ( 1)

 p( x)  x
dx
log(probability)
log(Prob > size)
log(rank)
Px

log(size)
0
=1
-1
1e3 samples from a
known distribution:
log10(P)
10
P( X  x) 
10  x
-2
-3
Px
-4
-1
0
1
10
10  x

2
x integer
3
4
log10(x)
5
=1
P( X  x)
Cumulative
Distributions
=0
Slope = -
Noncumulative
dP
Densities
=0
p( x)  
dx
=1
Slope
= -(+1)
=1
Correct
Cumulative
Distributions
Noncumulative
Densities
=0
=0
Wrong
The physics view
• Power laws are “suggestive of criticality”
• Self-organized criticality (SOC)
• Examples where this holds:
– Phase transitions in lab experiments
– Percolation models
– Rice pile experiments
• No convincing examples in technology, biology, ecology,
geophysical, or socio-economic systems
• Special case of “new science of complexity”
• Complexity “emerges” at a phase transition or
bifurcation “between order and disorder.”
• Doesn’t work outside the lab.
Data + Model/Theory
6
DC
5
WWW
4
3
2
1
SOC  = .15
Forest fire
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
SOC  = .15
Cumulative
distributions
Noncumulative
densities,
logarithmic
binning
 = .15
=.15
6
Web files
5
Codewords
4
Cumulative
Frequency
-1
3
Fires
2
-1/2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
Size of events
Log (base 10)
2
The HOT view of power laws
(w/ Jean Carlson, UCSB)
• The central limit theorem gives power laws as well as
Gaussians
• Many other mechanisms (eg multiplication noise) yield
power laws
• A model producing a power law is per se uninteresting
• A model should say much more, and lead to new
experiments and improved designs, policies, therapies,
treatments, etc.
The HOT view of power laws
• Engineers design (and evolution selects) for
systems with certain typical properties:
• Optimized for average (mean) behavior
• Optimizing the mean often (but not always)
yields high variance and heavy tails
• Power laws arise from heavy tails when there is
enough aggregate data
• One symptom of “robust, yet fragile”
HOT and fat tails?
• Surprisingly good explanation of statistics
(given the severity of the abstraction)
• But statistics are of secondary importance
• Not mere curve fitting, insights lead to new
designs
• Understanding  design
Examples of HOT fat tails?
•
•
•
•
•
•
Power outages
Detailed
Web/Internet file traffic
simulations
Forest fires
Commercial aviation delays/cancellations
Disk files, CPU utilization, …
Deaths or dollars lost due to man-made or natural
disasters?
• Financial market volatility?
• Ecosystem and specie extinction events?
• Other mechanisms, examples?
Examples with additional mechanisms?
•
•
•
•
•
•
•
Word rank (Zipf’s law)
Income and wealth of individuals and companies
Citations, papers
Social and professional networks
City sizes
Many others….
(Simon, Mandelbrot, …)
Data
6
DC
5
WWW
4
3
2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
Data + Model/Theory
6
DC
5
WWW
4
3
2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
6
Cumulative
Frequency
5
WWW files
Mbytes
4
(Crovella)
Most files
are small
(mice)
3
2
Most packets are
in large files
(elephants)
1
0
-1
-6
-5
Decimated data
Log (base 10)
-4
-3
-2
-1
0
1
Size of events
2
Router
queues
Mice
Sources
Network
Elephants
Router
queues
Mice
Delay sensitive
Sources
Network
Bandwidth
sensitive
Elephants
BW = Bandwidth sensitive traffic
Delay = Delay sensitive traffic
Log(bandwidth)
BW
cheap
Delay
Expensive
Log(delay)
• We’ll focus to begin with on similar tradeoffs in
internetworking between bandwidth and delay.
• We’ll assume TCP (via retransmission)
eliminates loss, and will return to this issue later.
Bulk transfers BW
(most packets)
Log(bandwidth)
Web navigation,
voice (most files)
Delay
Log(delay)
• Mice: many small files of few packets which
the user presumably wants ASAP
• Elephants: few large files of many packets for
which average bandwidth will be more
important than individual packet delay
• Most files are mice but most packets are in
elephants…
•…which is the manifestation of fat tails in the
web and internet.
Bulk transfers BW
(most packets)
Log(bandwidth)
Web navigation,
voice (most files)
Delay
Log(delay)
Claim I: Current
traffic dominated
by these two types
of flows
Claim II: Intrinsic
feature of many
future network
applications
Data
6
DC
5
WWW
4
3
2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
Data + Model/Theory
6
DC
5
WWW
4
3
2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
6
Cumulative
Frequency
5
WWW files
Mbytes
4
(Crovella)
Most files
are small
(mice)
3
2
Most packets are
in large files
(elephants)
1
0
-1
-6
-5
Decimated data
Log (base 10)
-4
-3
-2
-1
0
1
Size of events
2
6
Data
compression
WWW files
Mbytes
5
exponential
4
Cumulative
Frequency
All events are
close in size.
3
2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
Size of events
2
Based on frequencies of source word occurrences,
Select code
words.
To minimize message length.
Source coding for data compression
Source coding for data compression
Objectives:
• Optimally compress file
• Tractable compression
• Tractable decompression
Shannon:
• Optimally compress ensemble
• Tractable compression
• Tractable decompression
Kolmogorov:
• Optimally compress file
• Undecidable compression
• Intractable decompression
• Surprise: natural and practical
• Stochastic relaxation
• Philosophically important
• Turing, Godel, Chaitin, …
Shannon coding
Data
Compression
• Ignore value of information, consider only “surprise”
• Compress average codeword length (over stochastic
ensembles of source words rather than actual files)
• Constraint on codewords of unique decodability
• Equivalent to building barriers in a zero dimensional tree
• Optimal distribution (exponential) and optimal cost are:
length li  log( pi )
 pi  exp(cli )
Avg. length =
 pl
  pi log( pi )
i i
Shannon source coding
Minimize
expected
length
J
source words with
probabilities pi
 p l
i i
 r  1
i
length of
codewords li
 2 1
 ri  1  l   log(r )
i
i
 li
unique
decodability
Kraft’s inequality
Codewords
0
100
10
101
1
110
11
11100
1110
2
li
1
11101
11110
111
1111
0
100
101
110
11100
11101
11110
11111
11111
  2 li  1
 ri  1  l   log(r )
i
i
Kraft’s inequality =
Prefix-less code
Codewords
0
0 dimensional
(discrete) tree
100
10
101
1
110
11
11100
1110
2
li
1
11101
11110
111
1111
 cut in a 0-dim tree
  2 li  1
 ri  1  l   log(r )
i
i
0
100
101
110
11100
11101
11110
11111
11111
Kraft’s inequality =
Prefix-less code
Coding = building barriers
Source coding
2
li
Channel coding
1
Kraft’s inequality =
Prefix-less code
Channel noise
Control = building barriers
Minimize
J   pi li
 r  1
i
Leads to optimal solutions
for codeword lengths.
With optimal
cost
l (r )   log(r )
li   log( pi )
J   pi log( pi )
Equivalent to optimal barriers on a
discrete tree (zero dimensional).
J   pi li
 r  1
i
J   pi log( pi )
l (r )   log(r )
li   log( pi )
• Compressed files look like white noise.
• Compression improves robustness to limitations in
resources of bandwidth and memory.
• Compression makes everything else much more fragile:
– Loss or errors in compressed file
– Statistics of source file
• Information theory also addresses these issues at the
expense of (much) greater complexity
length li  log( pi )
 pi  exp(cli )
Data
6
5
How well does the
model predict the data?
DC
4
3
2
1
0
Avg. length =
 pl
  pi log( pi )
i i
-1
0
1
2
length li  log( pi )
 pi  exp(cli )
Data + Model
6
5
How well does the
model predict the data?
DC
4
3
Not surprising, because the
file was compressed using
Shannon theory.
2
1
0
Avg. length =
 pl
  pi log( pi )
i i
-1
0
1
2
Small discrepancy due to integer lengths.
Why is this a good model?
• Lots of models will reproduce an
exponential distribution
• Shannon source coding lets us
systematically produce optimal and
easily decodable compressed files
• Fitting the data is necessary but far
from sufficient for a good model
Web layout as generalized “source coding”
• Keep parts of Shannon abstraction:
– Minimize downloaded file size
– Averaged over an ensemble of user access
• Equivalent to building 0-dimensional
barriers in a 1- dimensional tree of
content
document
split into N files to
minimize download
time
A toy website model
(= 1-d grid HOT design)
Optimize 0-dimensional
cuts in a 1-dimensional
document
# links = # files
More complete
website models
(Zhu, Yu, Effros)
• Necessary for web layout design
• Statistics consistent with simpler models
• Improved protocol design (TCP)
• Commercial implications still unclear
Generalized “coding” problems
• Optimizing d-1 dimensional cuts in d dimensional
spaces…
• To minimize average size of files
• Models of greatly varying detail all give a consistent
story.
• Power laws have   1/d.
• Completely unlike criticality.
Web
Data compression
PLR optimization
Minimize
expected loss
J
 pili  ri  R
P: uncertain events
with probabilities pi
R: limited
resources ri
L: with
loss li
P
DC
source
WWW
user access
L
R
codewords decodability
files
web layout
document
split into N files to
minimize download
time
r = density of links or files
l = size of files
lr
1
d-dimensional
li = volume enclosed
ri = barrier density
li   , ri 
d
pi = Probability
of event

d 1
li
Resource/loss relationship:
lr
d
PLR optimization
J
=0
=1
=
“dimension”
 pili  ri  R
data compression
web layout
l (r ) 

r

c


1
 d
PLR optimization
J
=0
 =0 is
Shannon
source coding
 pili  ri  R
data compression

  log(r )

l (r )  
 c 
  r 1



 0
 0
Minimize average cost using
standard Lagrange multipliers
J 
 pili  ri
 R
Leads to optimal solutions for
resource allocations and the
relationship between the
event probabilities and sizes.
With optimal
cost
R  c 1
 pi  1
  log(r )

l (r )   c  
r 1
 


 0
 0





c  Rpi1 /(1  ) 
  1
li  
1 /(1  )
   p j



 j






  pi log(Rpi )   pi log( pi )

J 
1 
 c
1





 R  p 1    p 



i
i


 




 0
 0
Minimize average cost using
standard Lagrange multipliers
J 
 pili  ri
 R
Leads to optimal solutions for
resource allocations and the
relationship between the
event probabilities and sizes.
With optimal
cost
R  c 1
 pi  1
  log(r )

l (r )   c  
r 1
 


 0
 0





c  Rpi1 /(1  ) 
  1
li  
1 /(1  )
   p j



 j





  pi log( pi )
 0


1 
J  

1


1

   pi 1  
1   0

  


 




To compare with data.

c  Rpi1 /(1  ) 
  1
li  
1 /(1  )
   p j



 j




Forward
engineering
 pi 
ri 
Reverse
engineering
li 
To compare with data.
 pi 
ri 
pˆ i 
c1  li  c2 





c  Rpi1 /(1  ) 
  1
li  
1 /(1  )
   p j



 j




 (11/  )
Reverse
engineering
li 
plot
l , Pˆ 
i
sizes
from
data
i
compute
using
model
Cumulative





c  Rpi1 /(1  ) 
  1
li  
1 /(1  )
   p j



 j




Pˆi 
 pˆ  l
k i
pˆ i 
c1  li  c2 
 (11/  )
i
i
 li 1 
Data
6
DC
5
WWW
4
3
2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
Data + Model/Theory
6
DC
5
WWW
4
3
2
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
Typical web traffic
Heavy tailed
web traffic
 > 1.0
log(freq
> size)
p  s-
Is streamed out
on the net.
Web
servers
Creating fractal
Gaussian internet
traffic (Willinger,…)
log(file size)
3 
H
2
Fat tail
web traffic
time
Is streamed
onto the
Internet
creating long-range
correlations with
3 
H
2
Data + Model/Theory
6
DC
5
WWW
4
Are individual websites
3
distributed
like this?
2
1
Roughly, yes.
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
Data + Model/Theory
6
DC
5
WWW
4
How has the data
changed3 since 1995?
2
Steeper. Consistent
with more use of
cross hyperlinks.
1
0
-1
-6
-5
-4
-3
-2
-1
0
1
2
More complete
website models
(Zhu, Yu, Effros)
• More complex hyperlinks leads to steeper distributions
with 1<  < 2
• Optimize file sizes within a fixed topology:
• Tree:   1
• Random graph:   2
• No analytic solutions
The broader Shannon abstraction
• Information = surprise… and therefore ignoring
– Value or timeliness of information
– Topology of information
• Separate source and channel coding
– Data compression
– Error-correcting codes (expansion)
• Eliminate time and space
– Stochastic relaxation (ensembles)
– Asymptopia
• Brilliantly elegant and applicable, but brittle
• Better departure point than Kolmogorov, et al
What can we keep?
• Separation:
– Source and channel
– Congestion control and
error correction
– Estimation and control
• Tractable relaxations
– Stochastic embeddings
– Convex relaxations
What must we change?
• Add to information:
–
–
–
–
Value
Time and dynamics
Topology
Feedback
• More subtle treatment
of computational
complexity
• Naïve formulations
intractable
Log(bandwidth)
achievable
not
Distortion
Rate distortion theory
studies tradeoffs between
bandwidth and distortion
from lossy coding.
BW = Bandwidth sensitive traffic
Delay = Delay sensitive traffic
Log(bandwidth)
BW
cheap
Delay
Expensive
Log(delay)
• We’ll focus to begin with on similar tradeoffs in
internetworking between bandwidth and delay.
• We’ll assume TCP (via retransmission)
eliminates loss, and will return to this issue later.
Bulk transfers BW
(most packets)
Log(bandwidth)
Web navigation,
voice (most files)
Delay
Log(delay)
• Mice: many small files of few packets which
the user presumably wants ASAP
• Elephants: few large files of many packets for
which average bandwidth will be more
important than individual packet delay
• Most files are mice but most packets are in
elephants…
•…which is the manifestation of fat tails in the
web and internet.
Bulk transfers BW
(most packets)
Log(bandwidth)
Web navigation,
voice (most files)
Delay
Log(delay)
Claim I: Current
traffic dominated
by these two types
of flows
Claim II: Intrinsic
feature of many
future network
applications
Router
queues
Mice
Sources
Network
Elephants
Router
queues
Mice
Delay sensitive
Sources
Network
Bandwidth
sensitive
Elephants
Bulk transfers BW
(most packets)
Web navigation,
voice (most files)
Log(bandwidth)
Claim (channel): We can tweak TCP
using ECN and REM to make these flows
co-exist.
Delay
Log(delay)
Currently: Delays are
aggravated by queuing
delay and packet drops
from congestion caused
by BW traffic?
Specifically:
• Keep queues empty (ECN/REM).
• BW slightly improved (packet loss)
• Delay greatly improved (queuing)
• Provision network for BW
• “Free” QOS for Delay
• Network level stays simple
BW
Log(bandwidth)
The rare traffic
that can’t or won’t
will be expensive,
and essentially pay
for the rest.
Delay
Expensive
Log(delay)
Claim (source): Many (future)
applications are natural and
intrinsically coded into exactly this
kind of fat-tailed traffic.
BW
Delay
Expensive
Log(bandwidth)
Fat tailed traffic is
“intrinsic”
Log(delay)
• Two types of application traffic are important:
communications and control
• Communication to and/or from humans (from web to
virtual reality)
• Sensing and/or control of dynamical systems
• Claim: both can be naturally “coded” into fat-tailed
BW + delay traffic
• This claim needs more research
BW
Log(bandwidth)
Abstraction I
Delay
Expensive
Log(delay)
• Separate source and channel coding
• Source is coded into
– Delay sensitive mice
– Bandwidth sensitive elephants
• “Channel coding” = congestion control
Log(BW)
Loss?
Putting loss back
into the picture
Log(d)
• Packet loss can be handled by coding (application) or
retransmission (transport)
• Need coherent theory to perform tradeoffs
• Currently, congestion control and reliable transport are
intertwined
• What benefits would derive from some decoupling,
enabled by ECN or other explicit congestion control
strategies?
Optimization/control framework
• Application specific cost functions
J(app,delay,loss,BW) (assume to be minimized)
• Network resources:lines, routers, queues (energy,
spectrum, deployment, repair, stealth, security, etc)
• Comm/control network is embedded in other
networks (transportation, energy, military action, …)
• Robustness to uncertainties in users and resources
• Need to flesh out details for future scenarios
Optimization/control framework
• Global optimal allocation sets lower bound on
achievable performance
• Control problem is to find decentralized strategies
(eg TCP/IP) with (provably) near optimal
performance and robustness in dynamical setting
• Duality theory key to using network
• Coding and control interact in unfamiliar ways
• Naïve formulations intractable:
– Computation intractable
– Requires too much information not available to
decentralized agents
• Key is to find tractable relaxations
Optimization/control framework
• Pioneered by Kelly et al and extended by
Low et al.
• Ambitious goal: foundation for (much?)
more unified theory of computation, control,
and communications
• Hoped for outcome:
–
–
–
–
Rich theoretical framework
Motivated by practical problems
Yielding principled design of new protocols
And methods for deploying and managing
complex networks
Scalable Congestion Control
(Paganini, Doyle & Low ’01)
ROUTING + DELAY
x : source rates
R f ( s)
y : aggregate
link flows
LINKS
SOURCES
T
q : aggregate
prices per source
Rb ( s)
p : link prices
Robustness, evolvability/scalability, verifiability
Ideal
performance
Typical
design
IP
Robustness
Evolvability
Verifiability
Robustness of
HOT systems
Fragile
Robust
(to known and
designed-for
uncertainties)
Fragile
(to unknown
or rare
perturbations)
Robust
Uncertainties
Feedback and robustness
• Negative feedback is both the most powerful and most
dangerous mechanism for robustness.
• It is everywhere in engineering, but appears hidden as
long as it works.
• Biology seems to use it even more aggressively, but also
uses other familiar engineering strategies:
–
–
–
–
–
Positive feedback to create switches (digital systems)
Protocol stacks
Feedforward control
Randomized strategies
Coding
The Internet hourglass
Applications
Web
FTP
Mail
News
Video
Audio
ping
napster
Transport protocols
TCP SCTP UDP
ICMP
IP
Ethernet 802.11
Power lines ATM
Optical
Satellite Bluetooth
Link technologies
From Hari Balakrishnan
The Internet hourglass
Applications
Web
FTP
Mail
TCP
News
Video
Audio
ping
napster
Everything
Transport protocols
on IP
SCTP
UDP
ICMP
IP
Ethernet 802.11
IP on
Power lines ATM Optical
everything
Satellite Bluetooth
Link technologies
From Hari Balakrishnan
Consumers,
Applications
Applications
TCP/
IP
Hardware
Robust,
yet fragile
Robust
Mesoscale
Commodities,
Hardware
Uncertainty
Robust
Consumers,
Applications
Robust
Mesoscale
Uncertainty
Commodities,
Hardware
Consumers,
Applications
Yet fragile
Difficult
to change
Robust
Mesoscale
Commodities,
Hardware
Yet fragile
Protocols allow
for the creation
of large
complex
networks, with
rare but
catastrophic
cascading
failures.
Early computing
Various
functionality
Software
Digital
Hardware
Analog
substrate
Applications
Software
Modern
Computing
Operating
System
Hardware
Hardware
Varied
functionality
Robust
mesoscale
Uncertain
substrate
Various
functionality
Robust,
yet fragile
Digital
Analog
electronics
Consumers
Consumers
Barter
Money
Commodities
Commodities
Investors
Consumers
Markets,
Insitutions
Money
Investments
Commodities
The hourglass
Garments
Dress
Shirt
Slacks
Lingerie
Coat
Scarf
Sewing
Cloth
Wool
Cotton
Rayon
Polyester
Material technologies
Nylon
Tie
Consumers
Energy
•
•
•
•
110 V, 60 Hz AC
Gasoline
ATP, glucose, etc
Proton motive force
Energy
Producers
• Decentralized
• Asynchronous
Robust to:
• Network topology
• Application traffic
• Delays, link speeds
High performance
Applications
TCP/
IP
Hardware
Necessity:
Essentially only one
design is possible
• Decentralized
• Asynchronous
Robust to:
• Network topology
• Application traffic
• Delays, link speeds
Applications
TCP/
The existing
design
is incredible,IPbut…
Hardware
It’s a product of evolution,
and is not optimal.
High performance
Necessity:
Essentially only one
design is possible
Control
Theory
Computational
Information
Theory
All
design
Theory of
Complex systems?
Complexity
None
Statistical
Physics
Dynamical
Systems
1
dimension

Control
Theory
All
Computational
Biology
Information
Theory
design
• Non-equilibrium
• Highly tuned or optimized
• Finite but large dimension
Complexity
None
Statistical
Physics
Dynamical
Systems
1
dimension

Control
Theory
All
• Integrated horizontally and vertically
• Horizontal: control, communications,
computing
• Vertical: multiscale physics
design
None
Computational
Theory
needs
Information
Theory
• Status: nascent but promise results
Complexity
• Bonus: unexpected synergy
Statistical
Dynamical
Physics
Systems
1
dimension

Control
Theory
Computational
Information
Theory
All
design
• Ubiquity of power laws
• High shear turbulence
• Dissipation
• Quantum/classical transition
• Quantum measurement
Complexity
None
Statistical
Physics
Dynamical
Systems
1
dimension
