A Network Calculus for Multi-Hop Fading

Channels

Hussein Al-Zubaidy
Jörg Liebeherr
Almut Burchard

University of Toronto
 y ajt
Performance Analysis of Multihop Wireless Network
Through
traffic
Wireless link

Rx nN . . . nj Fading ch

odel
z Cross traffic Cross traffic Cross traffic
x y
Tx h Rx
Through
traffic
n z
1 . . . . . . N
Fading channel model Cross tra

Through
traffic
1

Intermediate nodes are store and Cross traffic

forward relays
Through
A fading channel is characterised traffic

model by its channel capacity Node model

Fading channel capacity

Fading channel capacity
Fading Channel Capacity

Channel capacity [Shannon 1948]

C(γ) = W log(1 + γ)

γ = γ̄|h|2 for fading channels

Channel gain h is a complex r.v.

Q: How do fading channel properties affect multihop

network performance?
 Network Model

Cross traffic Cross traffic Cross traffic Cross traffic

. . . nN A . . . D

Fluid-flow traffic, discrete time

fic
Arrival and service are independent
Cross traffic

I.i.d. cross traffic at each node

. . . N
Time-varying random service that is equal to the
Instantaneous channel capacity

C(γt ) = W log g(γt ) , γt = γ̄|ht |2

Computing this service distribution is hard!

Related Work: Multihop network performance analysis

Simplified channel models

FSMC model [Wang and Moayeri 1995][Sadeghi et al 2008]
more than two states models may not be tractable
not easily extended to multihop networks
ON-OFF model
tractable but very simplified model
used in queuing theory [Ishizaki 2007], network calculus
[Ciucu 2011], effective bandwidth [Hasan,Krunz,Matta 2004]
Effective capacity [Wu and Negi 2003]
log-MGF of the channel capacity
tractable only for low SNR where log(1 + γ) ' γ
Physical layer models [Hasna and Alouini 2003]
outage probability for AF wireless relay network
expression for MGF of end-to-end SNR
not suitable for network analysis
Network Calculus

A D
(min, +) dioid algebra S
Backlog: B(s) = A(0, s) − D(0, s)
Delay: W (s) = inf {u ≥ 0 : A(0, s) ≤ D(0, s + u)}

Dynamic server [Chang 2000]

A S
A(0, t) D
D(0, t) ≥ inf {A(0, u) + S(u, t)}
u≤t
delay = W (s) D(0, t)
=A ∗ S(0, t)
backlog = B(s)

Network service: s t

Snet (τ, t) = S1 ∗ S2 ∗ · · · ∗ SN (τ, t)

Network Analysis in Bit Domain

Bit domain A(t) S1 ... SN D(t)

Bit domain
Arrivals and departures are measured
log(X)
in bits
eX
For fading channels, service is given in terms of log(g(γt ))
Distribution of S is not easy to work with
SNR Domain

Transfer domain ...

S1 SN
(‘SNR domain’)

Bit domain A(t) S1 ... SN D(t)

Service in terms of g(γt ) rather than

X
log(X) log(g(γt )) – more
e
tractable
Qt−1
SNR service S(τ, t) = i=τ g(γi ) resides in the
SNR domain
SNR Domain

Transfer domain ...

(‘SNR domain’)
A(t) S1 SN D(t)

Bit domain A(t) S1 ... SN D(t)

Service in terms of g(γt ) rather than log(g(γt )) – more

tractable
Qt−1
SNR service S(τ, t) = i=τ g(γi ) resides in the
SNR domain
Our Approach

Transfer domain ...

(‘SNR domain’)
A(t) S1 SN D(t)

log(X)
eX

Bit domain A(t) S1 ... SN D(t)

SNR domain is governed by (min, ×) dioid algebra

Network SNR server

Snet (τ, t) = S1 ⊗ S2 ⊗ · · · ⊗ SN (τ, t)

 S
(min, ×) Network Calculus

Qt−1
Service: S(τ, t) = g(γi )
i=τ A S D
Qt−1
Arrival: A(τ, t) = i=τ eai

Departure: D(0, t) ≥ A ⊗ S(τ, t) = inf τ ≤u≤t A(τ, u)·S(u, t)
 
A(0,t)
Backlog: B(t) = log D(0,t)

Delay: W (t) = inf{u ≥ 0 : A(0, t) ≤ D(0, t + u)}

Computation of S1 ⊗ S2

Mellin transform: MX (s) = E[X s−1 ]

For two independent servers
t
X
MS1 ⊗S2 (s, τ, t) ≤ MS1 (s, τ, u) · MS2 (s, u, t)
u=τ

For N i.i.d. fading channels

N −1+t−τ
 

t−τ
MSnet (s, τ, t) ≤ · Mg(γ) (s) , ∀s < 1
t−τ

Moment bound: P r(X ≥ a) ≤ a−s MX (1 + s), ∀a, s > 0

Main Result: Statistical Performance Bounds
Define
min(τ,t)
X
M(s, τ, t) = MA (1 + s, u, t) · MS (1 − s, u, τ )
u=0

Backlog: P r B(t) > bε ≤ ε, where

n1
o
bε = inf log M(s, t, t) − log ε
s>0 s

Delay: P r W (t) > wε ≤ ε, where

n o
inf M(s, t + wε , t) ≤ ε
s>0
Cascade of N i.i.d. Rayleigh Channels

Service for Rayleigh channels

g(γ) = 1 + γ = 1 + γ̄|h|2
|h| ∼ Rayleigh r.v.
For i.i.d. Rayleigh fading channel
 t−τ
MS (s, τ, t) = e1/γ̄ γ̄ s−1 Γ(s, γ̄ −1 )

Arrivals: (σ(s), ρ(s)) bounded arrivals [Chang 2000]

MA (s, τ, t) ≤ e(s−1)·(ρ(s−1)·(t−τ )+σ(s−1)) , s>1

This traffic class includes Markov-modulated processes,

effective bandwidth, etc.
Performance Bounds of N Rayleigh Channels

Define:
4 1
V (s) = esρ(s) e1/γ̄ γ̄ −s Γ(1 − s, )
γ̄

Backlog: P r B(t) > bεnet ≤ ε, where

 
ε 1

bnet = inf σ(s) − N log(1 − V (s)) + log ε
s>0 s

Delay: P r W (t) > wε ≤ ε, where

( ε
)
es(−ρ(s)w +σ(s)) ε
· min 1, (V (s))w (wε )N −1

inf ≤ε
s>0 (1 − V (s))N
Numerical Results for N Rayleigh Channels

Model parameters
∆t = 1 ms
W = 20 kHz
(σ, ρ) bounded traffic
σ = 50 kb
ρ = 0 to 60 kbps
γ̄ = 0 to 40 dB
N = 1 to 100

We used deterministically bounded traffic, hence, the only

source of randomness is the fading channel!
Backlog Bounds for N Rayleigh Channels
& .

.<1*.1<7*
%'" .#!.1<7*A

:17!;<!*17.=,/>2<-.=<?17.6@=9
.$!.1<7*A
% .%!.1<7*A
."!.1<7*A
.#!!.1<7*A
bεnet vs. γ̄
$'"

$
ρ = 30 kbps #'"
ε = 10−4 #

!'"

!.
! " #! #" $! $" %! %" &!
()*+,-*./0,11*2.345.6789

$" -

--97/-798/
--!"-798/4
678!.9!/78-2+:1,9;-29<78-0=25

--$"-798/4
--%"-798/4
!#
--#"-798/4
!""-798/4
bεnet vs. ρ
!"
γ̄ = 10 dB
ε = 10−4
#

"-
!" !# $" $# %" %# &" &# #" ##
'(()*+,-(+./-012345
Backlog and Delays
!
#! -

-!-A-$!-B=<;C-:0)-0:6)
-!-A-%!-B=<;C-:0)-0:6)
#!
!# -!-A-$!-B=<;C-#!-0:6);

(i) ε(b) vs. γ̄ -!-A-%!-B=<;C-#!-0:6);

-!-A-$!-B=<;C-$!-0:6);

9:;;-<*:=+=>1>?@
!$ -!-A-%!-B=<;C-$!-0:6);
buffer size = 400kb #!

Waterfall curves for loss #!

!%

probability #!
!&

-
! " #! #" $! $" %!
'()*+,)-./+00)1-234-5678

!
#! .

.A!.B.#!.CD
!
.A .B.$!.CD
(ii) ε(w) vs. γ̄ #!
!%
.A!.B.%!.CD
.A!.B.E!.CD
:;<2,=;<1.>+<?,?;2;=@

N = 10 #!
!'

ρ = 20 kbps
!&
#!
Tighter delay bounds at
!#$

higher SNR #!

!#"
#! .
! " #! #" $! $" %! %"
()*+,-*./0,11*2.345.6789
Conclusions

New approach to analyze cascade of fading channels

Analysis in SNR domain using (min, ×) dioid algebra

Use Mellin transform and moment bound to compute

end-to-end bounds
Application to cascade of i.i.d. Rayleigh channels
Explicit bounds in terms of the physical channel parameters
Bounds scale linearly in N

(min, ×) dioid algebra has potential applications in models

with time varying channel models
Thank you
Q&A
Delay bounds

!
"! /

/!/=/&/+>?/-*./*-+.

(iii) ε(w) vs. EtoE delay /!/=/&/+>?/"!/*-+.7

/!/=/"!/+>?/-*./*-+.
/!/=/"!/+>?/"!/*-+.7
– ρ = 20 kbps

9:-01,:-*/;<-313:0:,2
"!
!$ /!/=/"&/+>?/-*./*-+.
/!/=/"&/+>?/"!/*-+.7

– Effect of N on the /!/=/#!/+>?/-*./*-+.

/!/=/#!/+>?/"!/*-+.7

violation prob. at low "!

!'

SNR is huge!
!(
"! /
! "! #! $! %! &! '!
)*+!,-!.*+/+.012/3-4*+/5678

#!! /

/=2+/2=8+
/#!/2=8+B
/$!/2=8+B
;28!<=!+28/8+3->/?=@28/7AB:

(! /%!/2=8+B

(iv) ε
wnet vs. γ̄ /"!/2=8+B

'!
– ρ = 30 kbps
– ε = 10−4 &!

$!
Fading Channels With Cross Traffic

Leftover SNR service: Ac (t) Dc (t)

S(τ, t)
So (τ, t) =
Ac (τ, t) S(τ, t)
Ao (t) Do (t)

Dynamic SNR server:

MSo (s, τ, t) = MS/Ac (s, τ, t) = MS (s, τ, t) · MAc (2 − s, τ, t)

N-node:
N −1+t−τ
 
(1−s)·N σc (1−s)
MSo,net (s, τ, t) ≤e
t−τ
(1−s)·ρc (1−s) t−τ


· Mg(γ) (s)e , s<1
Bounds of Rayleigh Channels With Cross Traffic

1 End-to-end Backlog of the through flow

( " #)
1
bεo,net (t) ≤ inf σo (s) + N σc (s) −


N log 1 − Vo (s) + log ε
s>0 s

2 Delay bound, we estimate for w ≥ 0

( )
es(−ρo (s)w+σo (s)+N σc (s))  w  N −1
inf · min 1, (Vo (s)) (w ) ≤
s>0 (1 − Vo (s))N

where,

Vo (s) = es·(ρo (s)+ρc (s) e1/γ̄ γ̄ −s Γ(1 − s, γ̄ −1 )

Numerical results
ε = 10−4 $# +
+"#+.94'>?+!,+@+#
W = 20 kHz +"#+.94'>?+!,+@+"#

7.4!89!'.4+:),;/9*+:9<.4+3=:6
+"#+.94'>?+!,+@+$#
∆t = 1 msec. "!
+"##+.94'>?+!,+@+#
+"##+.94'>?+!,+@+"#
(σ, ρ) bounded "#
+"##+.94'>?+!,+@+$#

through and cross

traffic !

σo = σc = 50 kb #+
! "# "! $# $!
%&'()*'+,-)..'/+012+3456

(i) bεo,net vs. γ̄

#! .

– ρo = 30 kbps ."!.8:905?.!;.@.!
."!.8:905?.!;.@."!
789!/:!089.3,;2-:<.3:=89.1>36

."!.8:905?.!;.@.#!
"&
."!!.8:905?.!;.@.!
."!!.8:905?.!;.@."!
."!!.8:905?.!;.@.#!
"!

(ii) bεo,net
vs. ρo
– γ̄ = 10 dB &

!.
! "! #! $! %! &! '!
())*+,-.),/0.123456
L. Le and E. Hossain.
Tandem queue models with applications to QoS routing in multihop wireless networks.
IEEE Trans. Mobile Comput., 7:1025–1040, 2008.

L. Le, A. Nguyen, and E. Hossain.

A tandem queue model for performance analysis in multihop wireless networks.
In Proc. IEEE WCNC, pages 2981–2985, March 2007.

N. Bisnik and A. A. Abouzeid.

Queuing network models for delay analysis of multihop wireless ad hoc networks.
Elsevier Ad Hoc Networks, 7(1):79–97, January 2009.

F. Kelly.
Notes on effective bandwidths.
In Stochastic Networks: Theory and Applications. (Editors: F.P. Kelly, S. Zachary and I.B.
Ziedins) Royal Statistical Society Lecture Notes Series, 4, pages 141–168. Oxford
University Press, 1996.

D. Wu and R. Negi.
Effective capacity: a wireless link model for support of quality of service.
IEEE Trans. Wireless Commun., 2(4):630–643, 2003.

C.-S. Chang.
Performance guarantees in communication networks.
Springer Verlag, 2000.

F. Ciucu.
Non-asymptotic capacity and delay analysis of mobile wireless networks.
In Proc. ACM Sigmetrics, pages 359–360, June 2011.

M. Fidler.
An end-to-end probabilistic network calculus with moment generating functions.
In Proc. IEEE IWQoS, pages 261–270, June 2006.

Y. Jiang and Y. Liu.

Stochastic network calculus.
Springer, 2008.

K. Mahmood, A. Rizk, and Y. Jiang.

On the flow-level delay of a spatial multiplexing MIMO wireless channel.
In Proc. IEEE ICC, pages 1–6, June 2011.

G. Verticale.
A closed-form expression for queuing delay in Rayleigh fading channels using stochastic
network calculus.
In Proc. ACM Q2SWinet ’09, pages 8–12, 2009.

G. Verticale and P. Giacomazzi.

An analytical expression for service curves of fading channels.
In Proc. IEEE Globecom, pages 635–640, Nov. 2009.

D. Wu and R. Negi.
Effective capacity: a wireless link model for support of quality of service.
IEEE Trans. Wireless Commun., 2(4):630–643, 2003.