Heterogeneous Congestion Control Protocols

Steven Low

CS, EEnetlab.CALTECH.edu

with A. Tang, J. Wang, D. Wei, CaltechM. Chiang, Princeton

Outline

� Review: homogeneous case� Motivating experiments� Model� Equilibrium

� Existence, uniqueness, local stability� Efficiency, fairness

� Slow timescale control

Tang, Wang, Low, Chiang. ToN, 2007Tang, Wang, Hegde, Low. Telecommunications Systems, Dec 2005Tang, Wei, Low, Chiang. ICNP, 2006

Bibliography!!!

�Bibliography!!!

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

Network model

for every RTT

{ W += 1 }

for every loss

{ W := W/2 }

Reno: (AI)

(MD)Jacobson

1989

=+

−=+

∑

∑

ililill

llli

i

ii

tptxRGtp

tpRxT

tx

)(),()1(

)(2

1)1(2

2

AI

MD

TailDrop

Network model: example

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

Duality model of TCP/AQM

� TCP/AQM

� Equilibrium (x*,p*) primal-dual optimal:

� F determines utility function U� G guarantees complementary slackness

� p* are Lagrange multipliers

) ,( ) ,(

***

***

RxpGpxpRFx T

==

cRxxU iix≤∑≥

subject to )( max0

Uniqueness of equilibrium� x* is unique when U is strictly concave� p* is unique when R has full row rank

Kelly, Maloo, Tan 1998Low, Lapsley 1999

Duality model of TCP/AQM

� TCP/AQM

� Equilibrium (x*,p*) primal-dual optimal:

� F determines utility function U� G guarantees complementary slackness

� p* are Lagrange multipliers

) ,( ) ,(

***

***

RxpGpxpRFx T

==

cRxxU iix≤∑≥

subject to )( max0

Kelly, Maloo, Tan 1998Low, Lapsley 1999

The underlying concave program also leads to simple dynamic behavior

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: HSTCP� α = 2 : Reno� α = : XCP (single link only) ∞ Low 2003

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

Low 2003

� α = 0: maximum throughput � α = 1: proportional fairness � α = 2: min delay fairness � α = : maxmin fairness ∞

Some implications� Equilibrium

� Always exists, unique if R is full rank� Bandwidth allocation independent of AQM or

arrival� Can predict macroscopic behavior of large scale

networks

� Counter-intuitive throughput behavior� Fair allocation is not always inefficient � Increasing link capacities do not always raise

aggregate throughput[Tang, Wang, Low, ToN 2006]

� FAST TCP� Design, analysis, experiments

[Jin, Wei, Low, ToN 2007]

Some implications� Equilibrium

� Always exists, unique if R is full rank� Bandwidth allocation independent of AQM or

arrival� Can predict macroscopic behavior of large scale

networks

� Counter-intuitive throughput behavior� Fair allocation is not always inefficient � Increasing link capacities do not always raise

aggregate throughput[Tang, Wang, Low, ToN 2006]

� FAST TCP� Design, analysis, experiments

[Jin, Wei, Low, ToN 2007]

Duality model

� Global stability in absence of feedback delay� Lyapunov function

� Kelly, Maulloo & Tan (1988)� Gradient projection

� Low & Lapsley (1999)� Singular perturbations

� Kunniyur & Srikant (2002)� Passivity approach

� Wen & Arcat (2004)

� Linear stability in presence of feedback delay� Nyquist criteria

� Paganini, Doyle, Low (2001), Vinnicombe (2002), Kunniyur& Srikant (2003)

� Global stability in presence of feedback delay� Lyapunov-Krasovskii, SoSTool

� Papachristodoulou (2005)� Global nonlinear invariance theory

� Ranjan, La & Abed (2004, delay-independent)

Outline

� Review: homogeneous case� Motivating experiments� Model� Equilibrium

� Existence, uniqueness, local stability� Efficiency, fairness

� Slow timescale control

The world is heterogeneous…

� Linux 2.6.13 allows users to choose congestion control algorithms

� Many protocol proposals� Loss- based: Reno and a large number of

variants� Delay- 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), …

Throughputs depend on AQM

� FAST and Reno share a single bottleneck router� NS2 simulation� Router: DropTail with variable buffer size � With 10% heavy-tailed noise traffic

FAST throughput

buffer size = 80 pkts buffer size = 400 pkts

Multiple equilibria: throughput depends on arrival

13M61Mpath 2

93M27Mpath 3

13M52MPath 1

eq 2eq 1

eq 1

eq 2

Tang, Wang, Hegde, Low, Telecom Systems, 2005

Dummynet experiment

13M61Mpath 2

93M27Mpath 3

13M52MPath 1

eq 2eq 1

Tang, Wang, Hegde, Low, Telecom Systems, 2005

eq 1

eq 2 eq 3 (unstable)

Dummynet experiment

Multiple equilibria: throughput depends on arrival

Multiple equilibria: single constraint sets

1 1

11x

21x

� Smallest example for multiple equilibria� Single constraint set but infinitely many equilibria� J=1: prices are non-unique but rates are unique� J>1: prices and rates are both non-unique

Some implications

dependentindependentbandwidthallocationon arrival

dependentindependentbandwidthallocation on AQM

non- uniqueuniqueequilibrium

heterogeneoushomogeneous

� Duality model:

, ***

= ∑l

illiii xpRFxcRxxU iix≤∑≥

s.t. )( max0

−+= ∑l

lliiii

iii pRx

TxF αγ

� Why can’t use Fi’s of FAST and Reno in duality model?

∑−=l

llii

ii pRx

TF

21 2

2

delay for FAST

loss for Reno

They use different prices!

� Duality model:

, ***

= ∑l

illiii xpRFxcRxxU iix≤∑≥

s.t. )( max0

−+= ∑l

lliiii

iii pRx

TxF αγ

� Why can’t use Fi’s of FAST and Reno in duality model?

∑−=l

llii

ii pRx

TF

21 2

2

They use different prices!

−= ∑i

lilil

l ctxRc

p )(1D

= ∑i

ililll txRtpgp )(),(D

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

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

Heterogeneous protocols

� Equilibrium: p that satisfies

∑

∑

>=≤

=

=

i,j ll

lji

jlil

ll

jlli

ji

ji

pcc

pxRpy

pmRfpx

0 if

)( : )(

)( )(

Duality model no longer applies !� pl can no longer serve as Lagrange multiplier

Heterogeneous protocols

� Equilibrium: p that satisfies

∑

∑

>=≤

=

=

i,j ll

lji

jlil

ll

jlli

ji

ji

pcc

pxRpy

pmRfpx

0 if

)( : )(

)( )(

Need to re-examine all issues� Equilibrium: exists? unique? efficient? fair?� Dynamics: stable? limit cycle? chaotic?� Practical networks: typical behavior? design guidelines?

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

−=

= ∑

))((

))(( ))((

γC

Notation

� Simpler notation: p is equilibrium ifon bottleneck links

� Jacobian:

� Linearized dual algorithm:

cpy )( =

)( : )( ppyp

∂∂=J

p(t)pp ∂=∂ )( *JγC

See Simsek, Ozdaglar, Acemoglu 2005for generalization

Outline

� Review: homogeneous case� Motivating experiments� Model� Equilibrium

� Existence, uniqueness, local stability� Efficiency, fairness

� Slow timescale control

Tang, Wang, Low, Chiang. Infocom 2005

Existence

Theorem

Equilibrium p exists, despite lack of underlying utility maximization

� Generally non-unique� There are networks with unique bottleneck

set but infinitely many equilibria� There are networks with multiple bottleneck

set each with a unique (but distinct) equilibrium

Regular networks

Definition

A regular network is a tuple (R, c, m, U) for which all equilibria p are locally unique, i.e.,

Theorem� Almost all networks are regular� A regular network has finitely many and

odd number of equilibria (e.g. 1)

0 )(det : )(det ≠∂∂= ppypJ

Regular networks

Proof idea:� Sard’s Theorem: critical value of a

continuously differentiable function over open set has measure zero

� Apply to y(p) = c on each bottleneck set � regularity

� Compact equilibrium set � finiteness

Regular networks

Proof idea:� Poincare- Hopf index theorem: if there

exists a vector field with non- singular dv/dp at every equilibrium and all trajectories move inward, then

� Dual algorithm defines such a vector field� Index theorem implies odd #equilibria

><−

=0det if 1 0det if 1

: )(index (p)(p)

pIJJ

L

p- pI )1( )(

eq

=∑

Global uniqueness

Corollary� If price mapping functions ml

j are linear and link-independent, then equilibrium is globally unique

e.g. a network of RED routers with slope inversely proportional to link capacity almost always has globally unique equilibrium

Theorem� If price heterogeneity is small, then equilibrium is

globally unique

0any for ]2,[

0any for ]2,[/1

/1

>∈

>∈jjLjj

l

llL

lj

l

aaamaaam

D

D

Global uniqueness

Remarks:

� Condition independent of U, R, c� Depends on m and size L of network � “Tight” from Index Theorem

Theorem� If price heterogeneity is small, then equilibrium is

globally unique

0any for ]2,[

0any for ]2,[/1

/1

>∈

>∈jjLjj

l

llL

lj

l

aaamaaam

D

D

Local stability:`uniqueness’ � stability

Theorem� If price heterogeneity is small, then the unique

equilibrium p is locally stable

0any for ]2,[

0any for ]2,[/1

/1

>∈

>∈jjLjj

l

llL

lj

l

aaamaaam

D

D

p(t)pp δγδ )( *J=DLinearized dual algorithm:

Equilibrium p is locally stable if

( ) 0 )( Re <pJλ

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−=∑

Outline

� Review: homogeneous case� Motivating experiments� Model� Equilibrium

� Existence, uniqueness, local stability� Efficiency, fairness

� Slow timescale control

Tang, Wang, Low, Chiang. Infocom 2005

Efficiency

Theorem

� Every equilibrium p* is Pareto efficient

Proof:

� Every equilibrium p* yields a (unique) rate x(p*)that solves

∑∑ ≤≥ j i

ji

ji

jix

cRxxUp t.s. )()(max *

0λ

Efficiency

Theorem

� Every equilibrium p* is Pareto efficient

� Measure of optimality

� Achieved:

∑∑ ≤=≥ j i

ji

jix

cRxxUV t.s. )(max : 0

*

∑∑=j i

ji

ji pxUpV ))(( : )( **

Efficiency

Theorem

� Every equilibrium p* is Pareto efficient� Loss of optimality:

� Measure of optimality

� Achieved:

∑∑ ≤=≥ j i

ji

jix

cRxxUV t.s. )(max : 0

*

∑∑=j i

ji

ji pxUpV ))(( : )( **

jl

jl

mm

VpV

D

D

max min )(

*

*

≥

Efficiency

Theorem

� Every equilibrium p* is Pareto efficient� Loss of optimality:

e.g. A network of RED routers with default parameters suffers no loss of optimality

jl

jl

mm

VpV

D

D

max min )(

*

*

≥

Intra-protocol fairness

Theorem� Fairness among flows within each type is

unaffected, i.e., still determined by their utility functions and Kelly’s problem with reduced link capacities

Proof idea:

� Each equilibrium p chooses a partition of link capacities among types, cj:= cj(p)

� Rates xj(p) then solve j

i

jjji

ji

xcxRxU

j≤∑

≥ t.s. )(max

0

Inter-protocol fairness

Theorem� Any fairness is achievable with a linear scaling of

utility functions

==

≤=

∑

∑≥

j

jj

i

jjji

ji

x

j

xaxX

cxRxUxj

: rates achievable all

t.s. )(maxarg : 0

Inter-protocol fairness

Theorem� Any fairness is achievable with a linear scaling of

utility functions� i.e. given any x in X, there exists µ s.t. an

equilibrium p has x(p) = x

( )llll

ll

jllij

i

ji

ji

ctpyp

tpmRftpx

−=

= ∑

))((

))((1 ))((

γµ

�

Outline

� Review: homogeneous case� Motivating experiments� Model� Equilibrium

� Existence, uniqueness, local stability� Efficiency, fairness

� Slow timescale control

Tang, Wang, Low, Chiang. Infocom 2005

Slow timescale control

Slow timescale scaling of utility function

∑

∑−+=+

=

ll

ll

jl

ji

ji

ji

ji

ji

jij

ij

i

tp

tpmt

ttqftx

)(

))(()1()( 1)(t

)()( )(

κµκµ

µ scaling of end--to-end price

slow timescale update of scaling factor

ns2 simulation: buffer=80pks

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

FAST throughput

without slow timescale control with slow timescale control

ns2 simulation: buffer=400pks

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

Summary: equilibrium structure

Uni-protocol�Unique bottleneck

set� Unique rates &

prices

Multi-protocol� Non-unique bottleneck

sets � Non-unique rates &

prices for each B.S.

� always odd� not all stable� uniqueness

conditions

Reno First

0 1000 2000 3000 4000 5000 6000 7000 80000

0.5

1

1.5

2

2.5

3

3.5

4

time(sec)

thro

ughput (p

kt / m

s)

FASTReno

0 1000 2000 3000 4000 5000 6000 7000 80000

0.5

1

1.5

2

2.5

3

3.5

4

time(sec)

thro

ughput (p

kt / m

s)

FASTReno

FAST First

0 1000 2000 3000 4000 5000 6000 7000 80000

0.5

1

1.5

2

2.5

3

3.5

4

time(sec)

thro

ughput (p

kt / m

s)

FASTReno

0 1000 2000 3000 4000 5000 6000 7000 80000

0.5

1

1.5

2

2.5

3

3.5

4

time(sec)

thro

ughput (p

kt / m

s)

FASTReno

Alpha trajectory

0 1000 2000 3000 4000 5000 6000 7000 80000

50

100

150

time(sec)

α (

pkt)

Fast FirstReno First

0 1000 2000 3000 4000 5000 6000 7000 80000

50

100

150

time(sec)

α (

pkt)

Fast FirstReno First