Markov Appro
Combinatorial Netw
Minghu
Joint wo
Soung‐Chang Liew, Ziyu Shao, Ca
oximation for
work Optimization
ua Chen
ork with
aihong Kai, and Shaoquan Zhang
Resource Alloca
ation is Critical
□ Utilize resource
– Efficiently
– Fairly
– Distributedly
□ A bottom‐up example
– Flow control (TCP)
2
Convex Networ
Popular an
□ Formulate resource alloc
maximization problem [K
□ Design distributed soluti
– Local decision, adapt to d
rk Optimization:
nd Effective
cation as a utility
Kelly 98, Low et. al. 99, …]
max X
Us(xs)
x≥0
s∈S
s.t.
Ax ≤ C
ions
dynamics
3
Example: Unde
max X 1
− Tr2xr
xr ≥0,r∈R Xr∈R
xs ≤ Cl, ∀
s.t.
s:l∈s
– End users: run TCP based
– Routers: drop packets ba
□ Local actions jointly solv
erstanding TCP
accessing efficiency + fairness
[Mo‐Walrand 00]
∀l ∈ L incoming rate less
than the link capacity
d on end‐to‐end measurement
ased on local information
ve global problem
4
Combinatorial Netw
Popular b
□ Joint routing and flow
x
?
□ Many others: Wireless
channel assignment, to
work Optimization:
but Hard
control problem
max X
Us(xs)
x≥0,A
s∈S
s.t. Ax ≤ C,
A ∈ {feasible
routing matrices}
s utility maximization,
opology control …
5
Observations a
Convex: solved
• Top‐down approach
• (mostly) convex problems
• Theory‐guided distributed
solutions
□ This talk: Explore theory‐gu
solutions for combinatorial
and Messages
Combinatorial: open
• Top‐down approach
• Combinatorial problems
• ??
uided design for distributed
l network problems
6
Markov Approximati
Convex network problems
Formulation
Penalty/decomposition
approach
Primal/dual/primal‐dual
design
ion: Our Perspective
Combinatorial network
problems
Formulation
Log‐sum‐exp approximation
Distributed Monte Carlo
Markov Chain
7
Existing Solutions
Optimi
□ Polynomial‐time approx
– Deterministic/randomize
– Centralized (in general)
□ Simulated annealing and
– No control of time‐comp
– Centralized (in general)
□ Our perspective: distribu
– Distributed simulated an
dynamics as special cases
for Combinatorial
ization
ximated solutions
ed solutions solutions
d Glauber dynamics
plexity tightly connect to
statistical physics
uted solution
nnealing and Glauber
s
8
Generic Form of Com
Optimizatio
maxf ∈F
□ System settings:
– A set of user configurat
– System performance un
□ Goal: maximize networ
choosing configuration
mbinatorial Network
on Problem
Wf .
tions, f =[f1, f2, …, f|R|]∈ F
nder f, Wf
rk‐performance by
ns
9
Exam
□ Wireless network utilit
– Configuration f: indepe
□ Channel assignments in
– Configuration f: one co
channel assignments
□ Path selection and flow
– Configuration f: one co
selected paths
□ Peering in Peer‐to‐Pee
mples
ty maximization New
endent set perspective
n WiFi networks
ombination of
w control New
perspective
ombination of
and new
er systems… solutions
10
Wireless Network U
max
z≥0,p≥
s.t.
□ zs: rate of user s
□ L: set of links, each with un
□ F : set of all independent s
□ pf: percentage of time f is a
Utility Maximization
x X Us(zs)
≥0 s∈SX
zs ≤ X pf , ∀l ∈ L
sX:l∈s,s∈S f :l∈f
pf = 1
f ∈F
Wireless link
nit capacity capacity constraints
sets (configurations)
active
11
Wireless Network U
max X Wirele
Us(zs) capacity c
z≥0,p≥0
s∈SX zs ≤ X pf ,
s.t.
sX:l∈s,s∈S f :l∈f
pf = 1
f ∈F
□ zs: rate of user s
□ L: set of links, each with
□ F : set of all independen
□ pf: percentage of time f i
Utility Maximization
ess link L3 ∅
constraints L1
L2
∀l ∈ L L2
L1 L3
3‐links interference L1L3
graph independent
unit capacity sets
nt sets (configurations)
is active
12
Scheduling Proble
XX
min max Us(zs) −
λ≥0 z≥0 sX∈S s∈
s.t. pf = 1.
f ∈F
□ An NP‐hard combinato
Independent Set probl
XX
max pf λl
p≥0 fX∈F l∈f
s.t. pf = 1.
f ∈F
em: Key Challenge
XX XX
zs λl + max pf λl
∈S l∈s p≥0 f ∈F l∈f
(scheduling)
orial Max Weighted
lem max X λl
= f ∈F
l∈f
13
Related Work
□ Wireless scheduling is NP‐h
□ It is recently shown that bo
scheduling problem approx
– [Wang‐Kar 05, Liew et. al. 0
Rajagopalan‐Shah 08, Liu‐Y
Srikant 09, …]
□ Our framework provides a
– Note that our framework a
problems
on Scheduling
hard [Lin‐Shroff‐Srikant 06, …]
ottom‐up CSMA can solve the
ximately
08, Jiang‐Walrand 08,
Yi‐Proutiere‐Chiang‐Poor 09, Ni‐
new top‐down perspective
applies to general combinatorial
14
Step 1: Log‐sum‐e
max X λl 1
β
f ∈F ≈
l∈f
max(x1, x2) ≈ 1 log (
10
□ Approximation gap:
□ The approximation be
approaches infinity
exp Approximation
⎛ ⎛ ⎞⎞
X X
1 log ⎝ exp ⎝β λl⎠⎠
β
f ∈F l∈f
(exp (10x1) + exp (10x2))
1 log |F|
β
ecomes exact as β
15
Step 1: Log‐sum‐e
max X λl 1
β
f ∈F ≈
l∈f
max XX max
pf λl
p≥0 p≥0
fX∈F l∈f
s.t.
pf = 1.
f ∈F
exp Approximation
⎛ ⎛ ⎞⎞
X X
1 log ⎝ exp ⎝β λl⎠⎠
β
f ∈F l∈f
Log‐sum‐exp is a concave and
closed function, double conjugate
is itself
X pf X − 1 X pf log pf
λl β
f ∈F Xl∈f f ∈F
s.t. pf = 1.
f ∈F
16
Big Picture After
□ The new max
z≥0,p≥
primal problem s.t.
□ Solution:
⎪⎪⎪⎧⎪⎪⎨⎪⎩zSλ˙˙scl h==edkαulslhehPUfs0s(f:loz∈rss),ps−f∈(SPβzλsl∈)−spλPelricl
r Approximation
≥0 X − 1X pf log pf
Us(zs) β
Xf ∈F
s∈SX pf , ∀l ∈ L
zs ≤
sX:l∈s,s∈R f :l∈f
pf = 1.
f∈F Distributed?
i+ TCP‐like ?
zs i+
l∈f pf (βλ) λl Local queue
centage of time.
17
Schedule by a Produ
X ⎛
max λl ≈ 1 log ⎝
β
f ∈F l∈f
pf (λ) pf
f ∈F
□ Computed by solving the Kar
the entropy‐approximated p
uct‐form Distribution
⎛ ⎛ ⎞⎞
XX
⎝ exp ⎝β λl⎠⎠
f ∈F l∈f
⎛⎞
X
f (λ) = 1 exp ⎝β λl⎠
C(βλ)
l∈f
f ∈F
rush‐Kuhn‐Tucker conditions to
problem
18
Step 2: Achieving p
⎛ X
pf (λ) = 1 exp ⎝β
C(βλ)
l∈
f ∈F
□ Rcleagsas rodf ptifm(λe)‐ raesv tehresi bstleeaM
– States: all the independe
– Transition rate: new desi
– Time‐reversible: detailed
pf(λ) Distributedly
⎞ f f0
X
λl⎠
∈f
f 000 f 00
pf (λ) qf,f0 = pf0 (λ) qf0,f
ady‐state distribution of a
Markov Chains
ent sets f ∈ F
ign space
d balance equation holds
19
Design Space: Two D
f f0 f f0
f 000 f 00 f 000 f 00
(a) (b)
pf (λ) qf,f0 =
□ 1) Add or remove transit
– Stay connected
– Steady state distribution
□ 2) Designing transition r
Degrees of Freedom
f f0 f f0
f 000 f 00 f 000 f 00
(c) (d)
= pf0 (λ) qf0,f
tion edge pairs
remains unchanged
rate
20
Design Goal: Distribu
Implement a Markov chain
=
Realize the transitions
□ What leads to distribut
– Every transition involve
– Transition rates Involve
uted Implementation
n L1
∅ L1L3
L2
L3
ted implementation?
es only one link
e only local Information
21
Every Transition Invo
□ From f to f’ = f ∪ {Li}: L
□ From f’ = f ∪ {Li} to f: L
L3
L2 L2
L1 L3 starts
3‐links conflict graph
olves Only One Link
Li starts to send
Li stops transmission
L1
∅ L1L3
s/stops
L3 L1 starts/stops
Designed Markov chain
22
Transition Rates Involve
□ Consider transition bet
□ λLi is the local queue le
□ Diffe³rePntq’s´ givedif
exp β l∈f λl ex
C(βλ) qf,f 0 =
⎛
X
exp ⎝ βλl −
l∈f 0
Only Local Information
tween f and f’ = f ∪ {Li}
ength of link Li
ffer³enPtimple´mentation
xp β l∈f 0 λl
qf 0,f
C(βλ)
⎞1
X
− βλl⎠ = exp βλLi
l∈f
23
Distributed Im
□ Link Li counts down at
exp(β λLi)
– Count down expires?
transmit
– Interference sensed? Fr
the count‐down, and
continue afterwards
□ Reinvent CSMA using a
top‐down approach
mplementation
rate L1
∅ L1L3
L2
reeze L3
L3
a L2
L1
24
The Overall Distr
⎨⎪⎪⎪⎧⎪⎪⎪⎩zλD˙˙slis==trikαblsuhhtPUeds0s(:Mlz∈ssC),s−M∈SCPzsal∈c−shλiPelvile+z∈ssfdpifst(
□ Converges to the optim
without time‐scale sep
– Proof utilizes Lyapunov
approximation, and mix
ributed Solution
Distributed?
TCP‐like
(βλ) i+
Local queue
λl
tribution pf (βλ). CSMA‐like
mal solution with or
paration
v arguments, stochastic
xing time bounds
25