# A Teletrafﬁc Theory for the Internet

` Thomas Bonald, Alexandre Proutiere
France Telecom R&D ´ ´ also afﬁliated with Ecole Normale Superieure

{thomas.bonald,alexandre.proutiere}@francetelecom.com
Tutorial of Performance 2005 October 2005

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 1

Teletrafﬁc theory

Born with the developing telephone network and exempliﬁed by the Erlang formula (1917):
B=
AC C! A2 2 + ...

1+A+

+

AC C!

where

B = blocking rate C = number of phone lines A = trafﬁc intensity in Erlangs

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 2

Teletrafﬁc theory

Born with the developing telephone network and exempliﬁed by the Erlang formula (1917):
B=
AC C! A2 2 + ...

1+A+

+

AC C!

where

B = blocking rate C = number of phone lines A = trafﬁc intensity in Erlangs

More generally, any capacity – demand – performance relationship

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 2

The insensitivity property

The Erlang formula does not depend on the distribution of call durations (beyond the mean)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 3

The insensitivity property
• •

The Erlang formula does not depend on the distribution of call durations (beyond the mean) It only requires Poisson call arrivals

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 3

The insensitivity property
• • •

The Erlang formula does not depend on the distribution of call durations (beyond the mean) It only requires Poisson call arrivals The key to simple and robust engineering rules

1917

2005

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 3

Flow-level modeling of the Internet

Proposed in 1998 by Massoulié & Roberts:
1 D= C −A

where

D = mean per-bit delay C = link capacity in bit/s A = trafﬁc intensity in bit/s

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 4

Flow-level modeling of the Internet

Proposed in 1998 by Massoulié & Roberts:
1 D= C −A

where

D = mean per-bit delay C = link capacity in bit/s A = trafﬁc intensity in bit/s

Based on fair sharing assumption (so-called processor-sharing model)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 4

Flow-level modeling of the Internet

Proposed in 1998 by Massoulié & Roberts:
1 D= C −A

where

D = mean per-bit delay C = link capacity in bit/s A = trafﬁc intensity in bit/s

• •

Based on fair sharing assumption (so-called processor-sharing model) Insensitive to all trafﬁc characteristics (beyond the trafﬁc intensity)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 4

Trafﬁc characteristics
flows think−times

Flows are generated within sessions

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 5

Trafﬁc characteristics
flows think−times

• •

Flows are generated within sessions Sessions typically arrive as a Poisson process

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 5

Trafﬁc characteristics
flows think−times

• • •

Flows are generated within sessions Sessions typically arrive as a Poisson process Deﬁnition of trafﬁc intensity − ﬂow arrival rate × mean ﬂow size (bit/s) − like telephone trafﬁc

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 5

Some key results
Loss networks (call blocking) Erlang, 1917 Engset, 1916 Bandwidth sharing (rate adaptation) Telatar & Gallager, 1995 Heyman et al, 1997 Roberts & Massoulié, 1998 Stamatelos & Koukoulidis, 1997 B & Virtamo, 2005

Gimpelson, 1965 Kaufman, 1981 Roberts, 1981 Brockmeyer et al, 1948 B & P, 2003 B, Massoulié, P & Virtamo, 2005 Kelly, 1986 Ross, 1995

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 6

Outline

Part 1: A single link − Processor-sharing model − Flow rate limits

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 7

Outline
Part 1: A single link − Processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 7

Outline
A brief reminder − The multiclass PS queue − Kelly networks − Whittle networks • Part 1: A single link − Processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 8

The multiclass PS queue
2 classes • Poisson arrivals of intensities λ1 , λ2
• • •

Exponential service requirements of parameters µ1 , µ2 PS service discipline
n1 n2 φ1 (n1 , n2 ) = , φ2 (n1 , n2 ) = n1 + n 2 n1 + n 2
n2

A reversible Markov process
(n1 + n2 )! λn1 λn2 1 2 π(n1 , n2 ) = π(0) n1 !n2 ! µn1 µn2 1 2

0

n1
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 9

Kelly networks
• • •

Multi-server PS (or symmetric) queues Poisson arrivals, exponential service requirements

Deterministic routes • A product-form distribution

(n1 + n2 )! λn1 λn2 ν m π(n1 , n2 , m) = π(0) n1 !n2 ! µn1 µn2 m! 1 2

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 10

Whittle networks
• •

PS queues with state-dependent service rates

Poisson arrivals, exponential service requirements n2 • Balance property
φ1 (n1 , n2 )φ2 (n1 − 1, n2 ) = φ1 (n1 , n2 − 1)φ2 (n1 , n2 ) 1 Φ(n1 , n2 ) = φ1 (n1 , n2 )φ2 (n1 − 1, n2 ) . . . φ1 (1, 0) λn 1 λn 2 π(n1 , n2 ) = π(0)Φ(n1 , n2 ) 1 1 2 2 µn µn 1 2
0

n1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 11

Outline
Part 1: A single link − Processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 12

The Erlang model
• • •

Poisson call arrivals of intensity λ Exponential call durations of mean τ An M/M/C/C queue
An π(n) = π(0) , n!

n≤C

where

A = λ × τ = trafﬁc intensity in Erlangs C = number of circuits

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 13

The Erlang model
• • •

Poisson call arrivals of intensity λ Exponential call durations of mean τ An M/M/C/C queue
An π(n) = π(0) , n!

n≤C

A = λ × τ = trafﬁc intensity in Erlangs C = number of circuits • The Erlang formula by PASTA B = π(C)=
AC C! A2 2 + ...

where

1+A+

+

AC C!

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 13

Insensitivity property
• •

Example: Erlang distribution with τ1 + τ2 = τ A Kelly queueing network

A n1 A n2 π(n1 , n2 ) = π(0) 1 2 n1 ! n2 ! n1 + n 2 ≤ C A1 + A 2 = A π(0) π(n1 , n2 ) = n! n =n n! An An1 An2 = π(0) n1 !n2 ! 1 2 n! =n

n1 +n2

1 +n2

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 14

Processor-sharing model
• • •

Poisson ﬂow arrivals of intensity λ Exponential ﬂow sizes of mean σ An M/M/1 queue π(n) = π(0)ρn ρ<1 where
A = λ × σ = trafﬁc intensity in bit/s C = capacity in bit/s ρ = A/C = link load

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 15

Mean per-bit delay
• D,

the ratio of the mean ﬂow duration to the mean ﬂow size

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 16

Mean per-bit delay
• D,

the ratio of the mean ﬂow duration to the mean ﬂow size • By Little’s law,
n = λ × σD ¯ ρ n= ¯ , 1−ρ A ρ= C =⇒ =⇒ n ¯ D= A 1 D= C −A

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 16

Mean per-bit delay
• D,

the ratio of the mean ﬂow duration to the mean ﬂow size • By Little’s law,
n = λ × σD ¯ ρ n= ¯ , 1−ρ

=⇒ =⇒

n ¯ D= A 1 D= C −A

A ρ= C

By insensitivity, the mean transfer delay of x bits is x × D

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 16

Insensitivity to the ﬂow size distribution
• •

Example 1: Erlang distribution with σ1 + σ2 = σ A Kelly network
(n1 + n2 )! n1 n2 π(n1 , n2 ) = π(0) ρ1 ρ2 n1 !n2 ! ρ1 + ρ 2 = ρ π(n1 , n2 ) = π(0)
n1 +n2 =n n1 +n2

n! n1 n2 ρ1 ρ2 = π(0)ρn n !n ! =n 1 2

ρ1 n1 = ¯ 1−ρ

ρ2 n2 = ¯ 1−ρ

1 D1 = D 2 = C −A
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 17

Insensitivity to the ﬂow arrival process
• •

Example 2: Two-ﬂow sessions with σ1 + σ2 = σ A Kelly network
(n1 + n2 )! n1 n2 ν m π(n1 , n2 , m) = π(0) ρ1 ρ2 n1 !n2 ! m! ρ1 + ρ 2 = ρ n! n1 n2 ρ1 ρ2 = π(0)eν ρn n !n ! =n 1 2 1 D1 = D 2 = C −A

π(n1 , n2 , m) = π(0)e
n1 +n2 =n

ν n1 +n2

ρ1 n1 = ¯ 1−ρ

ρ2 n2 = ¯ , 1−ρ

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 18

Flow throughput
• γ,

the inverse of the mean per-bit delay

γ =C −A

γ A C

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 19

Flow throughput
• γ, •

the inverse of the mean per-bit delay

The mean transfer delay of x bits is x/γ
γ =C −A

γ A C

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 19

Flow throughput
• γ, •

the inverse of the mean per-bit delay

The mean transfer delay of x bits is x/γ
1 0.8 Flow throughput 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 Link load 1 1.2 1.4

γ =C −A (C = 1)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 20

The Engset model
• K • • •

permanent sessions, jump-over blocking

Exponential call durations of mean τ Exponential think-time durations of mean ν −1 A closed Jackson network

K n π(n) = π(0) a , n

n≤C

where

a = ν × τ = per source virtual trafﬁc intensity C = number of circuits
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 21

Trafﬁc intensity
calls, τ think times, ν
−1

interarrival time

Effective per source trafﬁc intensity
τ a = −1 τ +ν a+1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 22

Trafﬁc intensity
calls, τ think times, ν
−1

interarrival time

Effective per source trafﬁc intensity
τ a = −1 τ +ν a+1

Overall trafﬁc intensity in Erlangs
a A=K× a+1
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 22

Insensitivity property
• •

Example: Erlang distribution with τ1 + τ2 = τ A closed Kelly network

K! an1 an2 1 2 π(n1 , n2 ) = π(0) (K − n)! n1 ! n2 ! n = n 1 + n2 ≤ C a1 + a 2 = a K π(n1 , n2 ) = π(0) n =n n! n1 n2 K n a1 a2 = π(0) a n !n ! n =n 1 2
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 23

n1 +n2

n1 +n2

The Engset formula

Number of ongoing calls seen by a new call
π 0 (n) ∝ π(n) × (K − n)ν K −1 n ∝ a , n≤C n

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 24

The Engset formula

Number of ongoing calls seen by a new call
π 0 (n) ∝ π(n) × (K − n)ν K −1 n ∝ a , n≤C n

Call blocking
B = π (C) =
0

(K − 1) . . . (K

aC − C) C!
C

1 + (K − 1)a + . . . + (K − 1) . . . (K − C) a C!

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 24

Engset vs. Erlang
• C = 20 • K = 25, 50, 250, ∞
1 Blocking probability 0.1 0.01 0.001 0.0001 1e-05 1e-06 0 0.2 0.4 0.6 0.8 1 Link load 1.2 1.4

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 25

The PS model with ﬁnite source
• K • • •

permanent sessions

Exponential ﬂow sizes of mean σ Exponential think-time durations of mean ν −1 A closed Jackson network

K! π(n) = π(0) (K − n)!

n

where

a = ν × σ = per source virtual trafﬁc intensity C = capacity in bit/s = a/C = per source virtual link load
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 26

Trafﬁc intensity
flows, σ think times, ν
−1

interarrival time

Effective per source trafﬁc intensity
σ C

σ = −1 +ν

a C

a +1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 27

Trafﬁc intensity
flows, σ think times, ν
−1

interarrival time

Effective per source trafﬁc intensity
σ C

σ = −1 +ν

a C

a +1

Overall trafﬁc intensity in bit/s
A=K×
a C

a +1
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 27

Mean per-bit delay
• D,

the ratio of the mean ﬂow duration to the mean ﬂow size

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 28

Mean per-bit delay
• D,

the ratio of the mean ﬂow duration to the mean ﬂow size • By Little’s law, n = λ × σD ¯
K

λ=
n=0

π(n)ν(K − n) = ν(K − n) ¯

=⇒

n 1 ¯ D= K −na ¯

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 28

Mean per-bit delay
• D,

the ratio of the mean ﬂow duration to the mean ﬂow size • By Little’s law, n = λ × σD ¯
K

λ=
n=0

π(n)ν(K − n) = ν(K − n) ¯

=⇒

n 1 ¯ D= K −na ¯

By insensitivity, the mean transfer delay of x bits is x × D

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 28

Insensitivity property
• •

Example: Erlang distribution with σ1 + σ2 = σ A closed Kelly network

K! n! π(n1 , n2 ) = π(0) (K − n)! n1 !n2 ! n1 + n 2 = n 1+ 2 = K! π(n1 , n2 ) = π(0) (K − n)! =n
n

n1 n2 1 2

,

n1 = ¯

1

n ¯

n2 = ¯

2

n ¯

n1 +n2

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 29

Finite vs. inﬁnite source
• C =1 • K = 10, 100, 1000, ∞
1 0.8 Flow throughput 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 Link load 1 1.2 1.4

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 30

Outline
Part 1: A single link − Processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 31

A common ﬂow rate limit
Maximum ﬂow bit rate c • Poisson ﬂow arrivals of intensity λ
• • •

Exponential ﬂow sizes of mean σ If C/c is an integer m, an M/M/m queue
π(n) = ρ<1 (ρm)n π(0) if n ≤ m n! n−m if n > m π(m)ρ

where

A = λ × σ = trafﬁc intensity in bit/s C = capacity in bit/s ρ = A/C = link load
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 32

Mean per-bit delay

Exact expression:
1 A B D= + × c C − (1 − B)A C − A

where B is the blocking probability in the corresponding Erlang model:
B=
Am m! A2 2 + Am m!

1+A+

... +

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 33

Mean per-bit delay

Exact expression:
1 A B D= + × c C − (1 − B)A C − A

where B is the blocking probability in the corresponding Erlang model:
B=

Am m! A2 2 + Am m!

1+A+

... +

Bound:

1 A 1 D≤ + × c C C −A
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 33

Flow throughput
• C =1 • c = 0.2, 0.4, 0.6, 0.8, 1
1 0.8 Flow throughput 0.6 0.4 0.2 0 0 0.2 0.4 0.6 Link load 0.8 1 Exact Bound

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 34

PS models with a common rate limit

Inﬁnite source (cf. Erlang model) − Poisson ﬂow/session arrivals − a multi-server PS queue

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 35

PS models with a common rate limit

Inﬁnite source (cf. Erlang model) − Poisson ﬂow/session arrivals − a multi-server PS queue

Finite source (cf. Engset model) − non-Poisson ﬂow arrivals − a closed network with one multi-server PS queue and one inﬁnite-server queue

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 35

Multirate loss systems
• K • • •

classes, bit rates c1 , . . . , cK

Poisson call arrivals of intensities λ1 , . . . , λK Exponential call durations of means τ1 , . . . , τK A reversible Markov process A n1 A n2 π(n1 , n2 ) = π(0) 1 2 n1 ! n2 ! n 1 c1 + n 2 c2 ≤ C where A1 = λ1 × τ1 = class-1 trafﬁc intensity in Erlangs A2 = λ2 × τ2 = class-2 trafﬁc intensity in Erlangs C = capacity in bit/s

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 36

Blocking probability

Class-1 blocking probability
A n1 A n2 1 2 n1 ! n2 !

B1 =

C−c1 <n1 c1 +n2 c2 ≤C A n1 1 n1 c1 +n2 c2 ≤C

A n2 2 n1 ! n2 !

Class-2 blocking probability
A n1 A n2 1 2 n1 ! n2 !

B2 =

C−c2 <n1 c1 +n2 c2 ≤C A n1 1 n1 c1 +n2 c2 ≤C

A n2 2 n1 ! n2 !
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 37

The Kaufman-Roberts formula

Assume C and c1 , c2 are integers

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 38

The Kaufman-Roberts formula
• •

Assume C and c1 , c2 are integers Deﬁne:
2 c1=1, c2=2

P (n) =
n1 c1 +n2 c2

A n1 A n2 1 2 n ! n2 ! =n 1

1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 38

The Kaufman-Roberts formula
• •

Assume C and c1 , c2 are integers Deﬁne:
2 c1=1, c2=2

P (n) =
n1 c1 +n2 c2

A n1 A n2 1 2 n ! n2 ! =n 1

1

Then:
1 P (n) = (A1 c1 P (n − c1 ) + A2 c2 P (n − c2 )) n with P (n) = 0 if n < 0
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 38

Example
• C = 100 • ck = 1, 5, 10, 30
1 Blocking probability 0.1 0.01 0.001 0.0001 1e-05 1e-06 0 0.2 0.4 0.6 0.8 1 Link load 1.2 1.4

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 39

Example (cont’d)
• C = 100 • ck = 1, 30
1 Blocking probability 0.1 0.01 0.001 0.0001 1e-05 1e-06 0 0.2 0.4 0.6 0.8 1 Link load 1.2 1.4

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 40

Multirate PS systems
• K • • •

classes, bit rates c1 , . . . , cK

Poisson ﬂow arrivals of intensities λ1 , . . . , λK Exponential ﬂow sizes of means σ1 , . . . , σK A Whittle network
π(n1 , n2 ) = π(0)Φ(n1 , n2 )An1 An2 1 2 A = A1 + A2 < C

where A1 = λ1 × σ1 = class-1 trafﬁc intensity in bit/s A2 = λ2 × σ2 = class-2 trafﬁc intensity in bit/s C = capacity in bit/s provided the allocation is balanced (so-called balanced fairness)
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 41

Balance property

The product of service rates φ1 , φ2 (allocated bit rates) does not depend on the considered path
1 Φ(n1 , n2 ) = φ1 (n1 , n2 )φ2 (n1 − 1, n2 ) . . . φ1 (1, 0)

n2

0

n1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 42

Balance property

The product of service rates φ1 , φ2 (allocated bit rates) does not depend on the considered path
1 Φ(n1 , n2 ) = φ1 (n1 , n2 )φ2 (n1 − 1, n2 ) . . . φ1 (1, 0)

n2

0

n1

A necessary and sufﬁcient condition for insensitivity!
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 42

Max-min fairness

Allocation by water-ﬁlling
φ1 (n1 , n2 ) = n1 c1 φ2 (n1 , n2 ) = n2 c2 φ1 (n1 , n2 ) = C − n1 c1 φ2 (n1 , n2 ) = n2 c2 φ1 (n1 , n2 ) = φ2 (n1 , n2 ) =
n1 n1 +n2 C n2 n1 +n2 C

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 43

Max-min fairness

Allocation by water-ﬁlling
φ1 (n1 , n2 ) = n1 c1 φ2 (n1 , n2 ) = n2 c2 φ1 (n1 , n2 ) = C − n1 c1 φ2 (n1 , n2 ) = n2 c2 φ1 (n1 , n2 ) = φ2 (n1 , n2 ) =
n1 n1 +n2 C n2 n1 +n2 C

The balance property is violated!

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 43

Balanced fairness

Allocation by balancing the service rates
Φ(n1 −1, n2 ) φ1 (n1 , n2 ) = , Φ(n1 , n2 ) Φ(n1 , n2 −1) φ2 (n1 , n2 ) = Φ(n1 , n2 )

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 44

Balanced fairness

Allocation by balancing the service rates
Φ(n1 −1, n2 ) φ1 (n1 , n2 ) = , Φ(n1 , n2 ) Φ(n1 , n2 −1) φ2 (n1 , n2 ) = Φ(n1 , n2 )

A unique balance function
1 1 Φ(n1 , n2 ) = n1 × n1 !c1 n2 !cn2 2

if n1 c1 + n2 c2 < C otherwise

1 Φ(n1 , n2 ) = (Φ(n1 −1, n2 ) + Φ(n1 , n2 −1)) C

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 44

Balanced fairness

Allocation by balancing the service rates
Φ(n1 −1, n2 ) φ1 (n1 , n2 ) = , Φ(n1 , n2 ) Φ(n1 , n2 −1) φ2 (n1 , n2 ) = Φ(n1 , n2 )

A unique balance function
1 1 Φ(n1 , n2 ) = n1 × n1 !c1 n2 !cn2 2

if n1 c1 + n2 c2 < C

1 Φ(n1 , n2 ) = (Φ(n1 −1, n2 ) + Φ(n1 , n2 −1)) otherwise C • The balance property is satisﬁed by construction

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 44

A recursive formula
• •

Assume C and c1 , c2 are integers Deﬁne:
2 c1=1, c2=2

P (n) =
n1 c1 +n2 c2 =n

Φ(n1 , n2 )An1 An2 1 2

1

1 Then: P (n) = (A1 c1 P (n − c1 ) + A2 c2 P (n − c2 )) n A1 A2 ¯≡ P (n) = P (n)+ P (n) P C −A C −A

n>C C−c1 <n≤C C−c2 <n≤C
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 45

Example
• C = 100 • ck = 1, 5, 10, 30
30 25 Flow throughput 20 15 10 5 0 0 0.2 0.4 0.6 0.8 1 Exact Bound

Link load
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 46

Comparison with max-min fairness
• C = 100 • ck = 1, 30
35 30 Flow throughput 25 20 15 10 5 0 0 0.2 0.4 0.6 0.8 1 Max-min fairness Balanced fairness

Link load
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 47

Outline
Part 1: A single link − The processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

We consider: − data networks only − no ﬂow rate limit

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 48

The linear network
2 1 C C 3

• Network state: n = (n1 , n2 , n3 ) numbers of active ﬂows on each

route
• Bandwidth allocation: φ(n) = (φ1 (n), φ2 (n), φ3 (n)) ∈ C
φ1

φ1 (n) + φ2 (n) ≤ C φ1 (n) + φ3 (n) ≤ C
φ2
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 49

φ3

The linear network
• A network of PS queues with state-dependent service rates - Each class-i ﬂow served at rate φi (n)/ni (TCP fairly shares the bandwidth among connections with the same characteristics) - A PS node per ﬂow class
2 1 C C 3 ρ1 φ1(n) ρ2 φ2(n) ρ3 φ3(n)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 50

The linear network
• Proportional fairness (Kelly’97)

n1 , φ1 (n) = n1 + n 2 + n 3 trafﬁc intensity on route i, ρi

φi (n) = C − φ1 (n), i = 2, 3

• Flow-level analysis with Poisson ﬂow arrivals on each route, - Stability: PF is stable if and only if ρ = (ρ1 , ρ2 , ρ3 ) ∈ C - Stationary distribution (C = 1): π(n) = π(0) - Flow throughput: 1 − ρ1 , γ1 = ρ ρ 1 + 1−ρ12+ρ2 + 1−ρ13+ρ3 γi = 1 − ρ1 − ρi , i = 2, 3
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 51

n1 + n 2 + n 3 n1 n2 n3 ρ1 ρ2 ρ3 n1

Outline
Part 1: A single link − The processor-sharing model − Flow rate limits • Part 2:Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 52

The model
• N ﬂow classes: ﬂows of the same class require the same

network resources (e.g., a set of links in wired network)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 53

The model
• N ﬂow classes: ﬂows of the same class require the same

network resources (e.g., a set of links in wired network)
• Network state: n = (n1 , . . . , nN ), ni number of class-i ﬂows

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 53

The model
• N ﬂow classes: ﬂows of the same class require the same

network resources (e.g., a set of links in wired network)
• Network state: n = (n1 , . . . , nN ), ni number of class-i ﬂows • Bandwidth allocation: φ(n) = (φ1 (n), . . . , φN (n)), class-i ﬂows

served at rate φi (n). C, a convex, compact, and monotone capacity set

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 53

The model
• N ﬂow classes: ﬂows of the same class require the same

network resources (e.g., a set of links in wired network)
• Network state: n = (n1 , . . . , nN ), ni number of class-i ﬂows • Bandwidth allocation: φ(n) = (φ1 (n), . . . , φN (n)), class-i ﬂows

served at rate φi (n). C, a convex, compact, and monotone capacity set
• Class-i ﬂows generated in sessions according to a Poisson

process, trafﬁc intensity ρi = λi /µi . ρ = (ρ1 , . . . , ρN )

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 53

The model
• N ﬂow classes: ﬂows of the same class require the same

network resources (e.g., a set of links in wired network)
• Network state: n = (n1 , . . . , nN ), ni number of class-i ﬂows • Bandwidth allocation: φ(n) = (φ1 (n), . . . , φN (n)), class-i ﬂows

served at rate φi (n). C, a convex, compact, and monotone capacity set
• Class-i ﬂows generated in sessions according to a Poisson

process, trafﬁc intensity ρi = λi /µi . ρ = (ρ1 , . . . , ρN )
• Issues:

- Is the network stable? - What is the mean time to transfer a class-i ﬂow?

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 53

Stability region
˘ ˘ • (i) If ρ ∈ C (C is the largest open subset of C), there exists a bandwidth allocation such that the network is stable
• (ii) If ρ ∈ C, the network is unstable under any bandwidth /

allocation
Proof. (i) If ρ

˘ ∈ C , let r ∈ C such that for all i, ri > ρi . The ﬁxed allocation φ(n) = r = {φ :
i

stabilizes the network. / (ii) If ρ ∈ C , there exists H

αi φi ≤ K} such that C ⊂ H and ρ ∈ H. Wi (t) workload of class-i ﬂows at time t, we have i αi Wi (t) → +∞ a.s.
φ2 H C ρ φ1
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 54

Utility-based allocations
• Usual allocation are based on the notion of utility

- Max throughput: U (r) = r

- Proportional fairness (Kelly’97): U (r) = ln r - Minimal potential delay (Massoulie-Roberts’99): U (r) = 1/r - α-bandwidth sharing (Mo-Walrand’00): U (r) = r 1−α /1 − α α=0 α→1 α=2 α → +∞ max throughput proportional fairness minimal potential delay max-min fairness (Rawls’71, Bertsekas-Gallager’87)
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 55

  max i ni U (φi (n)/ni )  φ(n) ∈ C

Utility-based allocations in practice
• Decentralized algorithms - e.g., A model of TCP: proportional fairness in wired networks can be arbitrarily closely approximated by the following decentralized algorithm, Kelly-Maullo-Tan’98: λi = φi /ni , ∂λi = wi − λi (t) ∂t pl (
l∈i j:l∈j

λj )

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 56

Utility-based allocations in practice
• Decentralized algorithms - e.g., A model of TCP: proportional fairness in wired networks can be arbitrarily closely approximated by the following decentralized algorithm, Kelly-Maullo-Tan’98: λi = φi /ni , ∂λi = wi − λi (t) ∂t pl (
l∈i j:l∈j

λj )

• Centralized algorithms - e.g. the gradient-based algorithm (for any capacity set, any utility function), Stolyar’05: at time t choose φ∗ ∈ C such that φ∗ = argmax
i

U (λi (t))φ∗ , i

λi (t + 1) = (1 − β)λi (t) + βφ∗ (t) i

The proportional fair algo (Tse) in CDMA/HDR system is built that way
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 56

Utility-based allocations: ﬂow-level stability
˘ • If ρ ∈ C, allocations based on utility functions of the form λ1−α /(1 − α), with α > 0, stabilize the network
Proof. Using classical ﬂuid limit and the following Lyapounov function

f (λ) =
i

λα+1 1/µi ρ−α i i α+1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 57

Utility-based allocations: ﬂow-level stability
˘ • If ρ ∈ C, allocations based on utility functions of the form λ1−α /(1 − α), with α > 0, stabilize the network
Proof. Using classical ﬂuid limit and the following Lyapounov function

f (λ) =
i

λα+1 1/µi ρ−α i i α+1

• The linear network paradox with the max throughput allocation,

B.-Massoulie’01
- Stability condition: ρ ∈ K = {ρ : ρ1 < (1 − ρ2 )(1 − ρ3 )} throughput in a dynamic scenario CMaximizing the throughput in all static scenarios can minimize the

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 57

Utility-based allocations: performance
• Proportional fairness on homogeneous linear, grid networks

(B.-Massoulie’01)
- On these networks, PF is insensitive - The stationary distribution is explicit

• For a general non trivial capacity set C, almost all utility-based

allocations are sensitive, e.g., maxmin is always sensitive

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 58

Utility-based allocations: performance
• Proportional fairness on homogeneous linear, grid networks

(B.-Massoulie’01)
- On these networks, PF is insensitive - The stationary distribution is explicit

• For a general non trivial capacity set C, almost all utility-based

allocations are sensitive, e.g., maxmin is always sensitive
• How can we predict the performance of these usual

allocations?
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 58

Balanced fairness
• Introduced by B.-P.’03 as the most efﬁcient insensitive

bandwidth allocation

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 59

Balanced fairness
• Introduced by B.-P.’03 as the most efﬁcient insensitive

bandwidth allocation
• Insensitivity implies the existence of a balance function Φ such

that ∀i,

Φ(n − ei ) φi (n) = Φ(n)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 59

Balanced fairness
• Introduced by B.-P.’03 as the most efﬁcient insensitive

bandwidth allocation
• Insensitivity implies the existence of a balance function Φ such

that ∀i,

Φ(n − ei ) φi (n) = Φ(n)

• Balanced fairness satisﬁes the capacity constraints

Φ(n − eN ) Φ(n − e1 ) ,..., Φ(n) Φ(n)

∈C

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 59

Balanced fairness
• Introduced by B.-P.’03 as the most efﬁcient insensitive

bandwidth allocation
• Insensitivity implies the existence of a balance function Φ such

that ∀i,

Φ(n − ei ) φi (n) = Φ(n)

• Balanced fairness satisﬁes the capacity constraints

Φ(n − eN ) Φ(n − e1 ) ,..., Φ(n) Φ(n)

∈C

• Efﬁciency means that φ(n) belongs to the border of C

Φ(0) = 1,

Φ(n − e1 ) Φ(n − eN ) ,..., ) ∈ C} Φ(n) = min{α : ( α α
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 59

Balanced fairness (cont’d)
• Invariant measure with balanced fairness

π(n) = π(0)Φ(n)ρn1 . . . ρnN 1 N

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 60

Balanced fairness (cont’d)
• Invariant measure with balanced fairness

π(n) = π(0)Φ(n)ρn1 . . . ρnN 1 N
• Network stability under balanced fairness if and only if

Φ(n)ρn1 . . . ρnN < +∞ 1 N
n

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 60

Balanced fairness (cont’d)
• Invariant measure with balanced fairness

π(n) = π(0)Φ(n)ρn1 . . . ρnN 1 N
• Network stability under balanced fairness if and only if

Φ(n)ρn1 . . . ρnN < +∞ 1 N
n

• Balanced fairness maximizes the probability the system is

empty

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 60

Balanced fairness (cont’d)
• Invariant measure with balanced fairness

π(n) = π(0)Φ(n)ρn1 . . . ρnN 1 N
• Network stability under balanced fairness if and only if

Φ(n)ρn1 . . . ρnN < +∞ 1 N
n

• Balanced fairness maximizes the probability the system is

empty
• The only possible Pareto-efﬁcient and insensitive allocation is

balanced fairness

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 60

Balanced fairness: stability
˘ • If ρ ∈ C, balanced fairness stabilizes the network ˜ Proof. Let Φ be a balance function satisfying the network capacity
constraints, i.e.,

˜ ˜ Φ(n − e1 ) Φ(n − eN ) ( ,..., ) ∈ C, ˜ ˜ Φ(n) Φ(n) ˜ ≤ Φ(n). ˜ Since (1 + )ρ ∈ C , the balance function Φ corresponding to the static allocation φ(n) = (1 + )ρ satisfy the network constraints and is stable.
then Φ(n) Finally,

Φ(n)ρn ≤
n n

˜ Φ(n)ρn < +∞

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 61

Balanced fairness: performance
• Under the stability condition, the performance can be evaluated

explicitly using Little’s formula ρi γi = Ei [ni ] and the network stationary distribution π(n) = π(0)Φ(n)ρn1 . . . ρnN 1 N

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 62

Outline
Part 1: A single link − The processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 63

Static routing
• A wired network is a set of L links and K routes where each

route k is a subset of links. Capacity of link l, Cl

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 64

Static routing
• A wired network is a set of L links and K routes where each

route k is a subset of links. Capacity of link l, Cl
• Static routing (N = K - the class of a ﬂow is deﬁned by a

route): capacity set C = {φ : φA ≤ C = (C1 , . . . , CL )} A is a N × L matrix, Akl = 1 if l ∈ k, Akl = 0 otherwise

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 64

Static routing
• A wired network is a set of L links and K routes where each

route k is a subset of links. Capacity of link l, Cl
• Static routing (N = K - the class of a ﬂow is deﬁned by a

route): capacity set C = {φ : φA ≤ C = (C1 , . . . , CL )} A is a N × L matrix, Akl = 1 if l ∈ k, Akl = 0 otherwise
• Stability condition: ∀l,
k:l∈k

ρk < C l

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 64

Static routing: balanced fairness
• An insensitive allocation is deﬁned by a balance function Φ

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 65

Static routing: balanced fairness
• An insensitive allocation is deﬁned by a balance function Φ • Capacity constraints

∀l,
k:l∈k

Φ(n − ek ) ≤ Cl Φ(n)

equivalent to ∀l, Φ(n) ≥ 1 Cl Φ(n − ek )
k:l∈k

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 65

Static routing: balanced fairness
• An insensitive allocation is deﬁned by a balance function Φ • Capacity constraints

∀l,
k:l∈k

Φ(n − ek ) ≤ Cl Φ(n)

equivalent to ∀l, Φ(n) ≥ 1 Cl Φ(n − ek )
k:l∈k

• Balanced fairness recursively deﬁned by Φ(0) = 1 and

Φ(n) = max
l

1 Cl

Φ(n − ek )
k:l∈k
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 65

The linear network
• A homogeneous 2-link line
1 C C 2 3

n2

+

n
+

n3 n1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 66

The linear network
• A homogeneous 2-link line • Both links are saturated:
2 1 C C 3

e.g. if n1 > 0 and n2 > 0, Φ(n) = Φ(n − e1 ) + Φ(n − e2 )

n2

+

n
+

n3 n1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 66

The linear network
• A homogeneous 2-link line • Both links are saturated:
1 C C 2 3

e.g. if n1 > 0 and n2 > 0, Φ(n) = Φ(n − e1 ) + Φ(n − e2 )
• Φ(n) is the number of direct
n2

+

n
+

paths from n to 0 Φ(n) = n1 + n 2 + n 3 n1
n3 n1

This is proportional fairness

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 66

A symmetric tree network
• A trunk and several branches with identical capacities
φ3
1

C

2

1

1

φ1

φ2

 C = (2, 1, 1, 1) A =  1 0 1  1 0 0

1 1 0

0

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 67

 0   1

A symmetric tree network (cont’d)
• Flow throughputs: γ1 = γ2 = γ3 =
(2− )(3− )(6+ ) , (4− )(9+ )

where

= ρ 1 + ρ2 + ρ3
1 0.8 Flow throughput 0.6 0.4 0.2 0 Max-min fair Proportional fair Balanced fair

0

0.5

1 1.5 Traffic intensity

2

• For max-min and proportional fairness, simulations with

Poisson ﬂow arrivals and exponentially distributed ﬂow sizes.
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 68

Recursive algorithm
• State space decomposition

Ω = N3 = {0} + Ω1 + Ω2 + Ω3 + Ω12 + Ω13 + Ω23 + Ω123 where ΩI = {n : ni > 0 iff i ∈ I}

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 69

Recursive algorithm (cont’d)
• Normalization constant G(ρ) =
n Φ(n)ρ n

G(ρ) = 1+G1 (ρ)+G2 (ρ)+G3 (ρ)+G12 (ρ)+G13 (ρ)+G23 (ρ)+G123 (ρ) where GI =
n∈ΩI

π(n)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 70

Recursive algorithm (cont’d)
• Normalization constant G(ρ) =
n Φ(n)ρ n

G(ρ) = 1+G1 (ρ)+G2 (ρ)+G3 (ρ)+G12 (ρ)+G13 (ρ)+G23 (ρ)+G123 (ρ) where GI =
n∈ΩI

π(n) Gij (ρ) =
ρi ρj (1−ρi )(1−ρj )

• We have Gi (ρ) =

ρi 1−ρi ,

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 70

Recursive algorithm (cont’d)
• Normalization constant G(ρ) =
n Φ(n)ρ n

G(ρ) = 1+G1 (ρ)+G2 (ρ)+G3 (ρ)+G12 (ρ)+G13 (ρ)+G23 (ρ)+G123 (ρ) where GI =
n∈ΩI

π(n) Gij (ρ) =
ρi ρj (1−ρi )(1−ρj )

• We have Gi (ρ) = • Recursion

ρi 1−ρi ,

ρ1 G23 (ρ) + ρ2 G13 (ρ) + ρ3 G12 (ρ) G123 (ρ) = 2−

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 70

Recursive algorithm (cont’d)
• Normalization constant G(ρ) =
n Φ(n)ρ n

G(ρ) = 1+G1 (ρ)+G2 (ρ)+G3 (ρ)+G12 (ρ)+G13 (ρ)+G23 (ρ)+G123 (ρ) where GI =
n∈ΩI

π(n) Gij (ρ) =
ρi ρj (1−ρi )(1−ρj )

• We have Gi (ρ) = • Recursion

ρi 1−ρi ,

ρ1 G23 (ρ) + ρ2 G13 (ρ) + ρ3 G12 (ρ) G123 (ρ) = 2−
• Flow throughtput

γi =

∂ ln G(ρ) ∂ρi

−1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 70

The store-and-forward bound
• Store-and-forward policy: the ﬂows on a given route are

transmitted sequentially on each link of this route Each link fairly shares its capacity among active ﬂows.

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 71

The store-and-forward bound
• Store-and-forward policy: the ﬂows on a given route are

transmitted sequentially on each link of this route Each link fairly shares its capacity among active ﬂows.
2 1 ρ1 C C 3 ρ2 C
PS

ρ3 C
PS

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 71

The store-and-forward bound
• Store-and-forward policy: the ﬂows on a given route are

transmitted sequentially on each link of this route Each link fairly shares its capacity among active ﬂows.
2 1 ρ1 C C 3 ρ2 C
PS

ρ3 C
PS

• The performance of store-and-forward provides a bound for

that of balanced fairness γi ≥
SF γi

=
l∈i

1 Cl −
k:l∈k

−1

ρk

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 71

Fast routing
• Each ﬂow class is assigned a set of routes

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 72

Fast routing
• Each ﬂow class is assigned a set of routes • All ﬂows of class i chooses one route in the subset s i at any

time. A set R of N × K matrices such that if R ∈ R, Rik = 1 if class-i ﬂows take route k and Rik = 0 otherwise

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 72

Fast routing
• Each ﬂow class is assigned a set of routes • All ﬂows of class i chooses one route in the subset s i at any

time. A set R of N × K matrices such that if R ∈ R, Rik = 1 if class-i ﬂows take route k and Rik = 0 otherwise
• At any time a routing scheme R ∈ R is chosen

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 72

Fast routing
• Each ﬂow class is assigned a set of routes • All ﬂows of class i chooses one route in the subset s i at any

time. A set R of N × K matrices such that if R ∈ R, Rik = 1 if class-i ﬂows take route k and Rik = 0 otherwise
• At any time a routing scheme R ∈ R is chosen • Capacity set

C = convex hull of {φ : ∃R ∈ R, φRA ≤ C}

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 72

Fast routing (cont’d)
• Example  1 0 0 1   1 0   1 0 

C = (1, 1),

A=

,

     1 0  or R =  0 1  R=    0 1 0 1
φ3

class 1 class 2 class 3 or

1

C

1

φ1

φ2
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 73

Trafﬁc splitting
• Each class i can use all routes in the set si at the same time

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 74

Trafﬁc splitting
• Each class i can use all routes in the set si at the same time • A set S of N × K stochastic matrices such that if S ∈ S,

Sik = 0 for all k except if k ∈ si
- Sik is the proportion of the total bandwidth φi (offered to class-i trafﬁc) class-i ﬂows get on route k

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 74

Trafﬁc splitting
• Each class i can use all routes in the set si at the same time • A set S of N × K stochastic matrices such that if S ∈ S,

Sik = 0 for all k except if k ∈ si
- Sik is the proportion of the total bandwidth φi (offered to class-i trafﬁc) class-i ﬂows get on route k • Capacity set

C = {φ : ∃S ∈ S, φSA ≤ C}

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 74

Trafﬁc splitting (cont’d)
• Example C = (1, 1),   0 1 0 0 1  
φ2
2

A= 1

,
C

 S= α  0
class 1 class 2 class 3 and

 1−α   1
1

1

1

1

1

φ1

φ3

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 75

Trafﬁc splitting (cont’d)
• Performance of balanced fairness

2(2 − )(3 − ) γ1 = γ 3 = , 12 − 5
Classes 1,3 2 1.5 1 0.5 0 Max-min fair Proportional fair Balanced fair 2 Flow throughput 1.5 1 0.5 0

γ2 = 2 −
Class 2 Max-min fair Proportional fair Balanced fair

0

0.5

1 1.5 Traffic intensity

2

0

0.5

1 1.5 Traffic intensity

2

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 76

Outline
Part 1: A single link − The processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 77

Downlink of cellular networks
• Formalism: a set of M transmission proﬁles, each

corresponding to a particular allocation of downlink radio resources

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 78

Downlink of cellular networks
• Formalism: a set of M transmission proﬁles, each

corresponding to a particular allocation of downlink radio resources
• At any time, a transmission proﬁle is chosen

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 78

Downlink of cellular networks
• Formalism: a set of M transmission proﬁles, each

corresponding to a particular allocation of downlink radio resources
• At any time, a transmission proﬁle is chosen • C is the M × N capacity matrix such that Cmi is the rate

allocated to class-i ﬂows in transmission proﬁle m

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 78

Downlink of cellular networks
• Formalism: a set of M transmission proﬁles, each

corresponding to a particular allocation of downlink radio resources
• At any time, a transmission proﬁle is chosen • C is the M × N capacity matrix such that Cmi is the rate

allocated to class-i ﬂows in transmission proﬁle m
• T the set of M -dimensional non-negative row vector summing

to 1. τ ∈ T corresponds to a particular schedule

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 78

A single cell
• We compare two access technologies - The ideal broadcast channel - TDMA access mode

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 79

A single cell
• We compare two access technologies - The ideal broadcast channel - TDMA access mode • For the sake of clarity: mobility/fading are not modelled

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 79

A single cell
• We compare two access technologies - The ideal broadcast channel - TDMA access mode • For the sake of clarity: mobility/fading are not modelled • User positions determine their feasible rate: when all resources

are allocated to class-i ﬂows, they receive a rate Ci

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 79

A single cell
• We compare two access technologies - The ideal broadcast channel - TDMA access mode • For the sake of clarity: mobility/fading are not modelled • User positions determine their feasible rate: when all resources

are allocated to class-i ﬂows, they receive a rate Ci

C2 C1

class 1

class 2

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 79

TDMA
• One user scheduled at a time: N transmission proﬁles (one per

class), C = diag(C1 , . . . , CN )
or
1

φ2

3

1

C

class 1

class 2

3

φ1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 80

TDMA
• One user scheduled at a time: N transmission proﬁles (one per

class), C = diag(C1 , . . . , CN )
or
1

φ2

3

1

C

class 1

class 2

3

φ1

• Scheduling - Fair time sharing: realized by proportional or balanced fairness φ1 (n) = 3n1 /(n1 + n2 ), φ2 (n) = n2 /(n1 + n2 ) - Fair rate sharing: realized by max-min fairness φ1 (n) = n1 /(n1 + 3n2 ), φ2 (n) = 3n2 /(n1 + 3n2 )
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 80

Performance of TDMA
• Fair time sharing γ1 = 3(1 − ), γ2 = 1 −

where

= ρ1 /3 + ρ2

• Fair rate sharing: A DPS queue, results by Fayolle et al’81
Class 1 3 2.5 2 1.5 1 0.5 0 0 0.25 0.5 0.75 1 Traffic intensity 1.25 1.5 Max-min fair Proportional fair Balanced fair 3 Flow throughput 2.5 2 1.5 1 0.5 0 0 0.25 0.5 0.75 1 Traffic intensity 1.25 1.5 Class 2 Max-min fair Proportional fair Balanced fair

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 81

The broadcast channel
• A Gaussian broadcast channel • Capacity set C = {φ : φ1 ≤ W log2 1 + P1 N1 , φ2 ≤ W log2 1 +
φ2
1

P2 N2 + P 1

, P1 +P2 ≤ P }

and 3 1

C

class 1

class 2

3

φ1

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 82

Performance of the broadcast channel
Class 1 3 2.5 2 1.5 1 0.5 0 0 0.25 0.5 0.75 1 1.25 Traffic intensity 1.5 1.75 Max-min fair Proportional fair Balanced fair 3 Flow throughput 2.5 2 1.5 1 0.5 0 0 0.25 0.5 0.75 1 1.25 Traffic intensity 1.5 1.75 Class 2 Max-min fair Proportional fair Balanced fair

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 83

Cell coordination
• 2 interfering base stations - 3 transmission proﬁles
φ2
3

3

(1)

class 1

class 2

2

C
3

(2)

class 1

class 2

2

3

φ1

2

2

(3)

class 1

class 2

3 0 C= 0 3  2 2
` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 84

Cell coordination: performance
3 2.5 Flow throughput 2 1.5 1 0.5 0 0 0.5 1 1.5 2 2.5 Traffic intensity 3 3.5 4 Max-min fair Proportional fair Balanced fair

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 85

Outline
Part 1: A single link − The processor-sharing model − Flow rate limits • Part 2: Networks − Bandwidth sharing − Application to wired, cellular, ad-hoc networks

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 86

Ad-hoc networks
• A set of L node-to-node links and K routes

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 87

Ad-hoc networks
• A set of L node-to-node links and K routes • A incidence matrix (i.e., Akl = 1l∈k )

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 87

Ad-hoc networks
• A set of L node-to-node links and K routes • A incidence matrix (i.e., Akl = 1l∈k ) • M transmission proﬁles used one at a time: C is the M × L

matrix, with Mml = capacity of link l in proﬁle m

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 87

Ad-hoc networks
• A set of L node-to-node links and K routes • A incidence matrix (i.e., Akl = 1l∈k ) • M transmission proﬁles used one at a time: C is the M × L

matrix, with Mml = capacity of link l in proﬁle m
• N = K classes of ﬂow

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 87

Example
class 2 class 1
(1) 2 2

class 3

3 links 2 transmission proﬁles

(2)

2

φ1
1

C
2

 A= 0 1  0 0
φ3

1 1

1

 0   1  

2

φ2

C=

2 0

0 2 2 0

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 88

Example (cont’d)
Class 1, 3 2 1.5 1 0.5 0 Max-min fair Proportional fair Balanced fair 2 Flow throughput 1.5 1 0.5 0 Class 2 Max-min fair Proportional fair Balanced fair

0

0.5

1 1.5 Traffic intensity

2

0

0.5

1 1.5 Traffic intensity

2

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 89

Summary
• Any data network can be represented by a network of

state-dependent PS queues

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 90

Summary
• Any data network can be represented by a network of

state-dependent PS queues
• Usual utility-based allocations - can be implemented in a centralized or distributed way - stabilize the network at ﬂow level - are sensitive (performance unknown)

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 90

Summary
• Any data network can be represented by a network of

state-dependent PS queues
• Usual utility-based allocations - can be implemented in a centralized or distributed way - stabilize the network at ﬂow level - are sensitive (performance unknown) • Balanced fairness (the most efﬁcient insensitive allocation) - stabilizes the network at ﬂow level - has an explicit performance that approximates that of usual utility-based allocations - A distributed algorithm to implement BF is not known yet

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 90

References
• Papers available at

http://perso.rd.francetelecom.fr/bonald http://perso.rd.francetelecom.fr/proutiere

` T. Bonald, A. Proutiere, A Teletrafﬁc Theory for the Internet – 91