See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322469467

Congestion control of data network by using anti-windup

approach

Article · January 2018

CITATION READS

1 236

3 authors:

Nabil El Fezazi El Houssaine Tissir

Sidi Mohamed Ben Abdellah University University Sidi Mohamed Ben Abdellah Faculty of Sciences
28 PUBLICATIONS 231 CITATIONS 143 PUBLICATIONS 1,267 CITATIONS

SEE PROFILE SEE PROFILE

El Haoussi Fatima
Sidi Mohamed Ben Abdellah University
45 PUBLICATIONS 462 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

singular system View project

Wind Tunnel View project

All content following this page was uploaded by El Haoussi Fatima on 18 January 2018.

The user has requested enhancement of the downloaded file.

International Journal of Ecological Economics and Statistics;
Volume. 39, Issue No. 1; Year 2018; Int. J. Ecol. Econ. Stat.;
ISSN 0973-1385 (Print), ISSN 0973-7537 (Online)
Copyright © 2018, [International Journal of Ecological Economics & Statistics]

Congestion control of data network by using

anti-windup approach

N. EL FEZAZI1 , F. EL HAOUSSI2 , E. H. TISSIR1 and T. ALVAREZ3

1
LESSI, Department of Physics
Faculty of Sciences Dhar El Mehraz
University Sidi Mohammed Ben Abdellah
BP 1796, Fes-Atlas, Morocco
fi[email protected]
[email protected]
2
IASCM, Department of Mathematic and Informatics
Polydisciplinary Faculty of Nador
BP 300, Selouane 62700, Nador, Morocco
elhaous [email protected]
3
Department of Systems Engineering and Automatic Control
University of Valladolid
47005 Valladolid, Spain
[email protected]

ABSTRACT

An Active Queue Management (AQM) scheme is design to control congestion in data net-
works, which includes anti-windup to deal with control signal saturation. More precisely, a
methodology is proposed to design advanced AQM systems capable of regulating queue
size even in the presence of significant disturbances. Hence, we first provide sufficient
conditions for stabilization for the equivalent class of systems, which are derived in terms
of LMI: this makes possible to derive optimization solutions that ensure performance and
stability for a large domain of initial conditions. This approach is validated with a numerical
example that illustrates the methodology, and the improvements with respect to previous
congestion control solutions.

Keywords: AQM, congestion control, anti-windup, signal saturation, stabilization.

Mathematics Subject Classification: 90B18, 34H05, 41A40, 93D15.

Journal of Economic literature (JEL) classification: C02, C60, C62, C63.

1 Introduction

Novel AQM algorithms are already playing a key role in data networks to cope with the in-
creasing user demands: Voice IP (VoIP) and video streaming are already pushing current
data networks to the congestion limits, as packet size and session duration vary significantly.

ISSN 0973-1385 (Print), ISSN 0973-7537 (Online)

www.ceser.in/ceserp
www.ceserp.com/cp-jour
 International Journal of Ecological Economics & Statistics

A significant research is being devoted to develop more efficient AQM techniques: The sim-
plest is Drop Tail, that drops packets arriving at a router when its buffer is full, which leads
to performance degradation due to excessive time-outs and restarts (Hollot, Misra, Towsley
and Gong, 2002). As an improvement, the Random Early Detection (RED) (Floyd and Ja-
cobson, 1993; Misra, Gong and Towsley, 2000) randomly drops/marks packets arriving at the
router before it is full to avoid busty traffics on the feedback signal (Hollot et al., 2002). However,
the low-pass filter used, which limits the closed-loop bandwidth (Hollot et al., 2002). Conse-
quently, variants of RED have been proposed for better congestion control: for example, (Ott,
Lakshman and Wong, 1999; Wang, Li, Hou, Sohraby and Lin, 2004). Random Exponential
Marking (REM) (Athuraliya, Low, Li and Yin, 2001) uses both the queue length and the input
rate as a congestion indication. In parallel, a fluid model of TCP dynamical behavior was de-
rived in (Misra et al., 2000), which makes possible to design controllers using traditional Control
Theory approaches, such as PI (Hollot, Misra, Towsley and Gong, 2001), PD (Sun, Chen, Ko,
Chan and Zukerman, 2003) or PID (Yanfie, Fengyuan and Chuang, 2003): this approach is
augmented here do deal with saturations in the control signals. Most of these techniques do
not take into account input saturation, but in AQM the control action is the discarding proba-
bility, frequently saturates, as it is a nonnegative number smaller than one. Accordingly, any
practical AQM should take into account this, as input saturations deteriorate the performance
and create instabilities.
Thus, this paper concentrates on the augmentation of congestion controllers with anti-windup
schemes, to deal with saturation due to inherent variations in traffic. Some anti-windup tech-
niques have been proposed in (Bender, 2013; El Fezazi, El Haoussi, Tissir and Tadeo, 2015;
El Fezazi, El Haoussi, Tissir, Husain and Zakaria, 2016; El Fezazi, Lamrabet, El Haoussi, Tis-
sir, Alvarez and Tadeo, 2016; Tarbouriech, Da Silva and Garcia, 2004; Tissir, 2014), where the
emphasis is on the transient performance caused by the saturation. The approach here is in-
spired by these previous results, but system discretization due to periodic sampling is explicitly
taken into account. Moreover, as stability during saturation is a central issue, the anti-windup
compensation is designed to enlarge the domain of initial conditions that mathematically en-
sure that the closed-loop system trajectories remain bounded. The mathematical objective is
then to design a controller capable of regulating the queue size at the router around a desired
value guaranteeing the stability, explicitly taking into account link capacity disturbances and
time-varying delay, using the discrete-time equivalent of the linearized TCP congestion window
model, and incorporating an anti-windup compensator. The proposed synthesis methodology
is based on Lyapunov functionals and LMI conditions that guarantee closed-loop stability of the
TCP/AQM system and minimization of the L2 −gain of the disturbance to the system output.
Then, the design of anti-windup controller is performed using LMIs.
It must be pointed out that the results are developed starting from a discrete-time state space
model of TCP/AQM based on the dynamic models developed by (Misra et al., 2000) of the
Transmission Control Protocol (TCP), which is derived using the assumptions that the data
traffic is equivalently represented by a fluid flow and the packet losses can be described by a
Poisson process. These assumptions have been shown to be valid in practice, as TCP has
been designed to be fair, and the packet losses are multiple.

38
 International Journal of Ecological Economics & Statistics

At the end of the paper, some simulation results will be presented to show the effectiveness of
the proposed method, and a comparison is made with other recent methods.

2 Problem Formulation

This section presents the problem and discusses some AQM models.

2.1 Dynamic Model of an AQM Router

In this paper, the network in Figure 1 is considered, with multiple server machines connect
to multiple client machines in a computer network. The network consists of n senders, n
receivers, and 1 bottleneck routers, which transports packets from senders to receivers. Large-
scale networks can be simplified as in Figure 1 in case of designing congestion controllers,
where one router is bottleneck in the network, running TCP flows.

Figure 1: Network topology used.

A model for this network was developed using fluid-flow and stochastic differential equation
analysis in (Misra et al., 2000)). Assuming that the AQM scheme implemented at the router
marks packets using Explicit Congestion Notification (ECN) (Yan, Gao and Ozbay, 2005) to
inform the TCP sources of impending congestion, and ignoring the TCP timeout mechanism,
the model that relates average values of the network variables is described by the following
coupled, nonlinear differential equations:

1 W (t)W (t − RT T (t))
Ẇ (t) = − p(t − RT T (t))
RT T (t) 2RT T (t − RT T (t))
N (t)
q̇(t) = −C(t) + W (t) (2.1)
RT T (t)

where
W (t) is the average TCP window size (packets);
q(t) is the average queue length (packets);
q(t)
RT T (t) is the round trip time = C(t) + Tp (secs);
C is the link capacity (packets/secs);

39
 International Journal of Ecological Economics & Statistics

Tp is the propagation delay (secs);

N is the number of sessions;
p ∈ [0 1] is the probability of packet marking/dropping.
As explained by (Misra et al., 2000), the first differential equation in (2.1) describes the TCP
window control dynamic and the second equation models the bottleneck queue length, from an
accumulated difference between packet arrival rate and link capacity. The congestion window
size W (t) increases by one every RTT when no congestion is detected, and is halved when
congestion is detected.
To derive the anti-windup mechanism (2.1) is linearized around the equilibrium point (that de-
pends on the nominal probability of packet marks). This nominal probability p0 fulfills p0 =
2N 2 2
(q0 +Tp C0 )2
, and at the equilibrium point I = (W0 = RT T C0
N , q0 = C0 (RT T − Tp ), p0 = W02
). In the
TCP/AQM network, each value of I is positive and the probability p0 is less than or equal to 1.
If we define δC = C − C0 with C = W, q, p, C, then, we can expressed the linearized version of
(2.1) as follows
−N   1  
δ Ẇ (t) = 2
δW (t) + δW (t − RT T (t)) − 2
δq(t) + δq(t − RT T (t))
RT T C0 RT T C0
RT T C02 RT T − Tp  
− 2
δp(t − RT T (t)) + 2
δC(t) + δC(t − RT T (t))
2N RT T C0
N 1 Tp
δ q̇(t) = δW (t) − δq(t) − δC(t)
RT T RT T RT T
δq(t)
RT T (t) = + RT T (2.2)
C0

2.2 TCP/AQM System Modelling in state space

Rewriting (2.2) in state space form yields

ẋ(t) = A0 x(t) + A1 x(t − τ (t)) + B0 u(t − τ (t)) + B1 w(t)

y(t) = Cyc x(t)
z(t) = Czc x(t) (2.3)

in which
     
−N −1 −N −1
δW (t) RT T 2 C0 RT T 2 C0 RT T 2 C0 RT T 2 C0
x(t) = , A0 = −1
, A1 = ,
δq(t) N
RT T RT T 0 0
     
−RT T C02 RT T −Tp RT T −Tp
2N 2 RT T 2 C0 RT T 2 C0 δC(t)
B0 = , B1 = −Tp , w(t) = ,
0 RT T 0 δC(t − RT T (t))
   
u(t) = δp(t), Cyc = , y(t) = δq(t), Czc = 1 , z(t) = RT T (t) − RT T
0 1 0 C0

where the states variables are then the congestion window and queue sizes, and the input
represents the marking probability.
As the AQM controller is by nature a discrete-time system, a discretized model of (2.3) assum-

40
 International Journal of Ecological Economics & Statistics

ing periodic sampling is

x(k + 1) = Ax(k) + Ad x(k − d(k)) + Bu(k − d(k)) + Bw w(k)

y(k) = Cy x(k)
z(k) = Cz x(k) (2.4)
T T T
where A = eA0 T , Ad = 0 eA0 s A1 ds, B = 0 eA0 s B0 ds, Bw = 0 eA0 s B1 ds, Cy = Cyc , and
Cz = Czc and d(k) is a positive integer representing the time delay of the system (that by
nature is time-varying) and satisfies dm ≤ d(k) ≤ dM where dm and dM are known positive
finite integers.
The disturbance vector w(k) is assumed to be limited in energy, that is, w(k) ∈ L2 . Hence for
some scalar δ, the bounds on the disturbance w(k) are the following:
∞
w(k)22 = wT (k)w(k) ≤ δ −1 < ∞ (2.5)
k=0

An anti-windup compensator is going to be proposed that shows graceful performance degra-

dation of the overall system in the presence of saturation. To this end, we reformulate the
problem into a state tracking problem, since the output is the combination of internal states.
In order to regulate (2.4) around a desired working point, we assume an controller, stabilizing
in absence of control bounds of the form:

xc (k + 1) = Ac xc (k) + Bc y(k)
yc (k) = Cc xc (k) + Dc y(k) (2.6)

As a consequence of the saturation, the interconnection with (2.4) is given by u(k) = sat(yc (k)),
where sat(yc (k)) = sign(yc (k))min{|yc (k)|, u0 }.
The following anti-windup compensator is proposed to mitigate the performance degradation
induced by the saturation, ensuring asymptotic stability of the closed-loop system:

xa (k + 1) = Aa xa (k) + Ba ψ(yc (k))

ya (k) = Ca xa (k) + Da ψ(yc (k)) (2.7)

Note that, ψ(yc (k)) corresponds to a decentralized dead-zone nonlinearity:

ψ(yc (k)) = yc (k) − sat(yc (k)) (2.8)

Considering the anti-windup compensator, the controller (2.6) is rewritten as follows

xc (k + 1) = Ac xc (k) + Bc uc (k) + ya (k)

yc (k) = Cc xc (k) + Dc uc (k) (2.9)

It follows that the augmented system corresponding to the closed-loop system (2.4)-(2.6)-(2.7)
can be represented by the following equation

ξ(k + 1) = Aξ(k) + Ad ξ(k − d(k)) − Bψ(Kξ(k − d(k))) + Bd ψ(Kξ(k)) + Bw w(k)

z(k) = Cz ξ(k) (2.10)

41
 International Journal of Ecological Economics & Statistics

where we define the following matrices:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
x(k) A 0 0 Ad + BDc Cy BCc 0
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
ξ(k) = ⎣ xc (k) ⎦ , A = ⎣ Bc Cy Ac Ca ⎦ , Ad = ⎣ 0 0 0 ⎦,
xa (k) 0 0 Aa 0 0 0
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
B 0 Bw    
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
B = ⎣ 0 ⎦ , Bd = ⎣ Da ⎦ , Bw = ⎣ 0 ⎦ , K = Dc Cy Cc 0 , Cz = Cz 0
0 Ba 0
As mentioned before, we aim to ensure a large set of initial states. For this, an estimate of
domain attraction will be used:
 
Ξ = φi (k), −dM ≤ k ≤ 0 : max φi (k) ≤ κ , i = 1, . . . , m

In mathematical terms, we are interested in the synthesis of the anti-windup compensator (2.7)
(i.e. in computing matrices Aa , Ba , Ca , and Da ), which ensures that the closed-loop trajectories
of the system remain bounded for any disturbance satisfying (2.5). Moreover, it should ensure
an upper bound for the L2 −gain between the disturbance w(k) and the regulated output z(k)
defined as follows (Chaibi, Tissir, Hmamed and Idrissi, 2013; El Fezazi, Tissir, El Haoussi,
Alvarez and Tadeo, 2017; El Haoussi and Tissir, 2007; Tissir, 2009)
∞ T
z(k)22 k=0 z (k)z(k)
Fzw 2∞ = 2 =  ∞ <γ (2.11)
w(k)2 k=0 w (k)w(k)
T

where Fzw is the closed-loop transfer function from w(k) to z(k). This ratio would be minimized
for a given set of expected network parameters.

3 Main Results

We first derive some delay-dependent conditions for stability of the TCP/AQM system with the
proposed anti-windup compensators. These conditions are given in terms of the existence of
solutions of some LMIs, based on Lyapunov functionals. This method provides a computable
criteria to check the stability for time-varying delays in the general case of dynamic system.
For the developments below the following is required:
For a matrix G we define the polyhedral set
 
S = ξ(k) ∈ n ; |(K(i) − G(i) )ξ(k)| ≤ u0(i)
The following lemma will be used later in this paper
Lemma 3.1. (Tarbouriech et al., 2004) If ξ(k) ∈ S, then the following relation is verified for any
diagonal positive matrix T
 
ψ T (Kξ(k))T ψ(Kξ(k)) − Gξ(k) ≤ 0

Remark 3.1. Lemma 3.1 allows a direct formulation of conditions in LMI form; moreover, the
obtained anti-windup synthesis conditions can be applied to stable or unstable systems, being
less conservative when the open-loop system is unstable.
Finally, for a positive scalar μ the trajectories of the system must not leave the set
 
ε(P, μ) = ξ(k) ∈ n ; ξ T (k)P ξ(k) ≤ μ−1

42
 International Journal of Ecological Economics & Statistics

3.1 Stability Results

Theorem 3.2. If there exists positive definite symmetric matrices P, Q,  R,

 and appropriately
         
sized matrices T1 , T2 , G1 , G2 , Y1 , Y2 , Y3 , Y4 , Y5 , Y6 such that the LMIs (3.1)-(3.3) are verified

⎡ ⎤
Π11 Π12 Π13 Π14 Π15 Π16 Π17 Π18
⎢ ⎥
⎢ ∗ Π22 Π23 Π24 Π25 Π26 Π27 0 ⎥⎥
⎢
⎢ ∗ ∗ Π33 Π34 Π35 Π36 0 0 ⎥⎥
⎢
⎢ ⎥
⎢ ∗ ∗ ∗ Π44 0 0 Π47 0 ⎥
⎢  < R,
⎥ < 0, Q  (3.1)
⎢ ∗ ∗ ∗ ∗ Π55 0 Π57 0 ⎥
⎢ ⎥
⎢ ⎥
⎢ ∗ ∗ ∗ ∗ ∗ Π66 Π67 0 ⎥⎥
⎢
⎢ ∗ ∗ ∗ ∗ ∗ ∗ Π77 0 ⎦⎥
⎣
∗ ∗ ∗ ∗ ∗ ∗ ∗ Π88

   
P PKT(i) − G
T
1(i)
T
P PKT(i) − G 2(i)
≥ 0, ≥ 0, (3.2)
∗ μu20(i) ∗ μu20(i)

μ − δ ≤ 0, (3.3)

where

Π11 = −P + Q
 + (dM − dm )R
 − Y1 − Y1T , Π12 = Y1 − Y2T , Π22 = −Q
 + Y2 + Y2T ,

Π13 = Y1 − Y3T , Π23 = Y2 + Y3T , Π33 = Y3 + Y3T , Π14 = −Y4T , Π24 = G
 T + Y T ,
1 4

Π34 = Y4T , Π44 = −2T1 , Π15 = G

 T − Y T , Π25 = Y T , Π35 = Y T , Π55 = −2T2 ,
2 5 5 5

Π16 = −Y6T , Π26 = Y6T , Π36 = Y6T , Π66 = −I, Π17 = PAT , Π27 = PATd ,
Π47 = −T1 BT , Π57 = T2 BTd , Π67 = BTw , Π77 = −P, Π18 = PCTz , Π88 = −γI

Then, there exists an anti-windup compensator (2.7) which ensures that the trajectories of the
system (2.10) converge asymptotically to the origin and are bounded for all initial conditions in
the ball
 
κ2 ≤ μ−1 − δ −1  P−1 )
λ(P−1 ) + (dM + dm )λ(P−1 Q
(dM − dm + 1)(dM + dm ) −1  −1 
+ λ(P RP ) (3.4)
2
with λ the maximal eigenvalue and κ = max φ.

Proof. Consider the following Lyapunov functional

V (k) = V1 (k) + V2 (k) + V3 (k)

k−1 −dm +1 k−1
= ξ T (k)P ξ(k) + ξ T (l)Qξ(l) + ξ T (m)Rξ(m) (3.5)
l=k−d(k) l=−dM +2 m=k+l−1

43
 International Journal of Ecological Economics & Statistics

Computing the difference of the Lyapunov functional gives

ΔV1 (k) = ξ T (k + 1)P ξ(k + 1) − ξ T (k)P ξ(k), (3.6)

k k−1
ΔV2 (k) = ξ T (l)Qξ(l) − ξ T (l)Qξ(l)
l=k+1−d(k+1) l=k−d(k)
k−1
= ξ T (k)Qξ(k) − ξ T (k − d(k))Qξ(k − d(k)) + ξ T (l)Qξ(l)
l=k+1−dm
k−1 k−dm
− ξ T (l)Qξ(l) + ξ T (l)Qξ(l), (3.7)
l=k+1−d(k) l=k+1−d(k+1)
−dm +1
 k k−1

ΔV3 (k) = ξ (m)Rξ(m) −
T
ξ (m)Rξ(m)
T

l=−dM +2 m=k+l m=k+l−1

k−dm
= (dM − dm )ξ T (k)Rξ(k) − ξ T (l)Rξ(l) (3.8)
l=k+1−dM

As ∀Q < R, one can easily see that

k−1 k−1
− ξ T (l)Qξ(l) ≤ − ξ T (l)Qξ(l), (3.9)
l=k+1−d(k) l=k+1−dm

k−dm k−dm
ξ T (l)Qξ(l) ≤ ξ T (l)Rξ(l) (3.10)
l=k+1−d(k+1) l=k+1−dM

Then, from (2.10) and (3.6)-(3.10), it follows that

 T 
ΔV (k) ≤ Aξ(k) + Ad ξ(k − d(k)) + Bd ψ(Kξ(k)) − Bψ(Kξ(k − d(k))) + Bw w(k) P Aξ(k)

+Ad ξ(k − d(k)) + Bd ψ(Kξ(k)) + Bw w(k) − Bψ(Kξ(k − d(k)))
+ξ T (k)(−P + (dM − dm )R + Q)ξ(k) − ξ T (k − d(k))Qξ(k − d(k)) (3.11)

Using the Newton-Leibniz formula, for any appropriately dimensioned matrices Y1,...,6 the fol-
lowing holds:

 k−1
ξ T (k)Y1 + ξ T (k − d(k))Y2 + y T (j)Y3 + ψ T (Kξ(k − d(k)))Y4
j=k−d(k)
 k−1 
+ψ T (Kξ(k))Y5 + wT (k)Y6 − ξ(k) + ξ(k − d(k)) + y(j) = 0 (3.12)
j=k−d(k)

where y(j) = ξ(j + 1) − ξ(j).

Then, applying Lemma 3.1 and taking into account (3.11)-(3.12), the following inequality holds
1 T
ΔV (k) − wT (k)w(k) + z (k)z(k) ≤ η T (k)(Υ + LT P L)η(k) (3.13)
γ

44
 International Journal of Ecological Economics & Statistics

where
⎡ ⎤ ⎡ ⎤T ⎡ ⎤
Υ11 Υ12 Υ13 Υ14 Υ15 Υ16 AT ξ(k)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ Υ22 Υ23 Υ24 Υ25 Υ26 ⎥ ⎢ ATd ⎥ ⎢ ⎥
⎢ ∗ ⎥ ⎢ ⎥ ⎢ ξ(k − d(k)) ⎥
⎢ ∗ ∗ Υ33 Υ34 Υ35 Υ36 ⎥ ⎢ 0 ⎥ ⎢ k−1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ j=k−d(k) y(j) ⎥
Υ=⎢ ⎥, L = ⎢ ⎥ , η(k) = ⎢ ⎥,
⎢ ∗ ∗ ∗ Υ44 0 0 ⎥ ⎢ −B ⎥T ⎢ ψ(Kξ(k − d(k))) ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ∗ ∗ ∗ ∗ Υ55 0 ⎥ ⎢ BTd ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ψ(Kξ(k)) ⎦
∗ ∗ ∗ ∗ ∗ Υ66 Bw T
w(k)

and
1 T
Υ11 = −P + Q + (dM − dm )R − Y1 − Y1T + C Cz , Υ12 = Y1 − Y2T ,
γ z
Υ22 = −Q + Y2 + Y2T , Υ13 = Y1 − Y3T , Υ23 = Y2 + Y3T , Υ33 = Y3 + Y3T ,
Υ14 = −Y4T , Υ24 = GT1 T1T + Y4T , Υ34 = Y4T , Υ44 = −2T1 , Υ15 = GT2 T2T − Y5T ,
Υ25 = Υ35 = Y5T , Υ55 = −2T2 , Υ16 = −Y6T , Υ26 = Υ36 = Y6T , Υ66 = −I

Then, it is clear that if

Υ + LT P L < 0 (3.14)

then
1 T
ΔV (k) − wT (k)w(k) + z (k)z(k) < 0 (3.15)
γ

Accordingly, the following condition is obtained by applying the Schur complement to (3.14)
⎡ ⎤
Υ11 Υ12 Υ13 Υ14 Υ15 Υ16 Υ17
⎢ ⎥
⎢ ∗ Υ22 Υ23 Υ24 Υ25 Υ26 Υ27 ⎥
⎢ ⎥
⎢ ∗ ∗ Υ33 Υ34 Υ35 Υ36 0 ⎥
⎢ ⎥
⎢ ⎥
⎢ ∗ ∗ ∗ Υ44 0 0 Υ47 ⎥ < 0, (3.16)
⎢ ⎥
⎢ ∗ ∗ ∗ ∗ Υ55 0 Υ57 ⎥
⎢ ⎥
⎢ ⎥
⎣ ∗ ∗ ∗ ∗ ∗ Υ66 Υ67 ⎦
∗ ∗ ∗ ∗ ∗ ∗ Υ77

where

Υ17 = AT P, Υ27 = ATd P, Υ47 = −BT P, Υ57 = BTd P, Υ67 = BTw P, Υ77 = −P

Pre- and post-multiplying (3.16) by Δ = diag{P −1 , P −1 , P −1 , T1−1 , T2−1 , I, P −1 }, applying the

Schur complement and taking the following changes of variables

P = P −1 , Ω
 = PΩP, Ω = Q, R, Y1 , Y2 , Y3 , Λ
 = ΛP, Λ = G1 , G2 , Y6 ,

T1 = T1−1 , T2 = T2−1 , Y4 = T1 Y4 P, Y5 = T2 Y5 P.

we obtain the inequality (3.1) of Theorem 3.2.

45
 International Journal of Ecological Economics & Statistics

Since (3.1) holds, the condition (3.15) is satisfied. Now, summing up (3.15) from 0 to ∞ with
respect to k yields
∞  1 
V (∞) < V (0) + wT (k)w(k) − z T (k)z(k) (3.17)
γ
k=0

Under the zero initial condition V (0) = 0 and by noting that V (∞) ≥ 0, we have (2.11) which
implies that system (2.10) has its restricted L2 −gain from w(k) to z(k) less than γ.
The LMIs in (3.1) and (3.2) ensure that the trajectories are contained inside the ellipsoid ε(P, μ),
∀k, once ε(P, μ) ⊂ S. This is verified by the following conditions
   
P KT(i) − GT1(i) P KT(i) − GT2(i)
≥ 0, ≥0
∗ μu20(i) ∗ μu20(i)

These matrices give the ellipsoidal inclusion LMIs (3.2) by pre- and post-multiplied by Δ =
diag{P, I}. Moreover, from the Lyapunov functional (3.5), it follows that
 dM + dm 
V (0) ≤ λ(P ) + (dM + dm )λ(Q) + (dM − dm + 1) λ(R) φ2 = β (3.18)
2
Then, we have

ξ T (k)P ξ(k) ≤ V (k) ≤ V (0) + w(k)22 ≤ β + δ −1 ≤ μ−1

Hence, for all k the trajectories of the system do not leave the set ε(P, μ), concluding the
proof.

We focus now on guaranteing the stability of the following system, which can be seen as a
particular case of system (2.10).

x(k + 1) = Ax(k) + Ad x(k − d(k)) + Bu(k) + Bw w(k)

y(k) = Cy x(k)
z(k) = Cz x(k) (3.19)

Then, the augmented system is given by

ξ(k + 1) = Aξ(k) + Ad ξ(k − d(k)) + (Bd − B)ψ(Kξ(k)) + Bw w(k)

z(k) = Cz ξ(k) (3.20)

The following corollary gives a condition to stabilize system (3.19)

Corollary 3.3. If there exists positive definite symmetric matrices P, Q,  R,

 and appropriately
      
sized matrices T2 , G2 , Y1 , Y2 , Y3 , Y5 , Y6 such that the LMIs (3.21)-(3.22) are verified
⎡ ⎤
Π11 Π12 Π13 Π15 Π16 Π17 Π18
⎢ ⎥
⎢ ∗ Π22 Π23 Π25 Π26 Π27 0 ⎥
⎢ ⎥
⎢ ∗ ∗ Π33 Π35 Π36 0 0 ⎥
⎢ ⎥
⎢ ⎥  < R,

⎢ ∗ ∗ ∗ Π55 0 Π57 − T2 BT 0 ⎥ < 0, Q (3.21)
⎢ ⎥
⎢ ∗ ∗ ∗ ∗ Π66 Π67 0 ⎥
⎢ ⎥
⎢ ⎥
⎣ ∗ ∗ ∗ ∗ ∗ Π77 0 ⎦
∗ ∗ ∗ ∗ ∗ ∗ Π88

46
 International Journal of Ecological Economics & Statistics

 
T
P PKT(i) − G 2(i)
≥ 0, μ − δ ≤ 0 (3.22)
∗ μu20(i)

Then, there exists an anti-windup compensator as defined in (2.7) which ensures that the
trajectories of the system (3.20) converge asymptotically to the origin and are bounded for
every initial condition in the following ball
 
κ2 ≤ μ−1 − δ −1  P−1 )
λ(P−1 ) + (dM + dm )λ(P−1 Q
(dM − dm + 1)(dM + dm ) −1  −1 
+ λ(P RP ) (3.23)
2
with κ = max φ.

Remark 3.2. In deriving Theorem 3.2, the slack variable Y1,...,6 were introduced in order to
offer additional degrees of freedom for the optimization. It can be seen from the above Proof
that ΔV (k) remains unaffected by equation (3.12), so these matrices lead to more flexible LMI
conditions in (3.1), reducing possible conservatism in Theorem 3.2 and the subsequent results,
as will be shown in the numerical examples.

3.2 Implementation Constraints

The control signal for the network is given by

⎡ ⎤
δW (k)
 ⎢ ⎥
⎢ δq(k) ⎥ 
u(k) = sat(Kξ(k)) = sat K 1 K2 K3 K4 ⎢ ⎥ (3.24)
⎢ x (k) ⎥
⎣ c ⎦
xa (k)

As the used AQM are based on state feedback, and having in mind that the state δW (k) =
W (k) − W0 , is not directly available at routers in real networks, it is necessary to take this
constraint into account. Some authors have proposed to use an observer to estimate this
state (Chen, Hung, Liao and Yan, 2007; Manfredi, Di Bernardo and Garofalo, 2009) whereas
others expressed W (k) − W0 as a function of the rate mismatch at routers, which can be
estimated (Zhang, Ye, Ma, Chen and Li, 2007). The approach in this paper uses the following
approximation:

RT T  N W (k)  RT T  
W (k) − W0 = − C0 = f low rate − C0 (3.25)
N RT T N
Furthermore, as in (Athuraliya et al., 2001), it must be pointed out that the rate of mismatch
is the rate at which the queue length grows when the buffer is nonempty. Therefore, we can
δq 1
approximate it by T where Tis the sampling frequency. Hence, (3.24) becomes
  
δp(k) = sat 0 K1 RT
N T + K2 K3 K4
T ξ(k)

Finally, it must be pointed out that to implement our AQM controller (3.1), we first discretize
(2.3).

47
 International Journal of Ecological Economics & Statistics

3.3 Anti-windup Optimization

3.3.1 Minimization of γ

As (Bender, 2013), in order to minimize γ where μ−1 = δ −1 a solution should be given for the
following problem where the initial condition is null

min γ
subject to (3.1) − (3.3) (3.26)

3.3.2 Maximization of κ

Now, we consider the free-disturbance case (w(k) = 0). In practice the bounds on the time-
varying delay dM and dm can be known. In order to ensure the stability of system (2.10) by
using the Theorem 3.2, the admissible initial conditions must verify (3.4). Note that the smaller
the maximal eigenvalues of P−1 , P−1 Q
 P−1 , and P−1 R
P−1 , the larger κ for which (3.4) is
verified. Hence, the problem of finding Aa , Ba , Ca , and Da is transformed into the maximization
of the region of stability that can be achieved by minimizing these maximal eigenvalues. With
this aim, consider the following auxiliary LMIs as in (El Fezazi, El Haoussi, Tissir, Alvarez and
Tadeo, 2017; El Haoussi, Tissir and Tadeo, 2014) where P = P−1 , Q =Q  −1 , and R=R −1
     
σ1 I I σ2 I P σ3 I P
≥ 0, ≥ 0, ≥0 (3.27)
I P P Q  P R 

Consequently, the condition (3.4) implies that

  dM − dm + 1  
κ2 σ1 + (dM + dm ) σ2 + σ3 ≤ μ−1 (3.28)
2
Then, we construct a feasibility problem as follows
 
min tr PP + Q
Q+R
R
     
P ∗ 
Q ∗ 
R ∗
subject to (3.1)-(3.3),(3.27),(3.28), ≥ 0, ≥ 0, ≥ 0 (3.29)
I P I 
Q I 
R

Based on the above conditions, the proposed controller can be designed for given dM and dm
by using the following cone complementarity algorithm:
   
Step 3.1. Choose a small δ and set P, P, Q,
 Q,
 R,
 R,
 σ1 , σ2 , σ3 = P, P, Q,
0
 Q,
 R,
 R,
 σ1 , σ2 , σ3
that satisfies the constrained minimization (3.29). Then, fix Δ where δ = δ + Δ.
Step 3.2. Solve the following LMI minimization problem in the matrix variables P, P, Q,
 Q,
 R,


and R
 
min tr PP0 + Q
Q0 + R
R0 + P0 P + Q
0 Q
+R
0 R


subject to LMIs in (3.29)

Step 3.3. Substitute the new matrix variables into (3.29). If the result is feasible, then set
δ = δ + Δ and repeat Step 2; otherwise, δ = δ − Δ is the desired estimate: Stop.

48
 International Journal of Ecological Economics & Statistics

4 Simulations

In order to demonstrate the effectiveness and applicability of proposed design methodology,

two examples are derived, with simulations provided to compare with existing works. The first
one aims at studying the conservativeness. In the second example, some Matlab simulations
are provided to compare the proposed controller with previous controllers.
Example 4.1. Consider the closed-loop system (3.20) with parameters for which control values
are saturated at ±10 where
       
0.8 0 −0.1 −0.1 1 1 0
A= , Ad = ,B = , Cy = ,
0 0.97 0 −0.1 0 0 1
   
0.0718 0.0389 −0.0213 0.0001
Ac = , Bc = ,
−0.0502 −0.0012 0.0621 0.0080
   
Cc = 0.0184 0.0213 , Dc = −0.0228 −0.0087

Using the cone complementarity algorithm, the results obtained in Corollary 3.3 ensures the
asymptotic stability; the estimated domain of attraction for different delay ranges are shown in
Table 1, which can be compared with the results in (Negi, Purwar and Kar, 2012).

Table 1: Values of κ obtained for several delay ranges.

Method Reference (Negi et al., 2012) This paper
1 ≤ d(k) ≤ 3 0.5151 2.3
1 ≤ d(k) ≤ 4 0.3852 1.5
1 ≤ d(k) ≤ 5 0.2918 1.1
1 ≤ d(k) ≤ 6 Inf easible 0.9

It is clear that the obtained stability radius κ is significantly larger than those obtained
 in(Negi
0.7
et al., 2012). The corresponding matrices of (2.7) are Aa = 0.3, Ba = 0.3, Ca = , and
0.7
Da = 0.5.
Example 4.2. In this section we illustrate our methodology based on a TCP/IP router queue
model borrowed from (Bender, 2013), given in the form of (2.3). The nominal parameters are
RT T = 0.2467, C0 = 3750, q0 = 175, and N = 60. Then, we can deduce that W0 = RT T C0
N p0 =
,
2
W02
, and Tp = RT T − C0 . Finally, the matrices of (2.6) are Ac = 0, Bc = 1, Cc = 8.4969 × 10−6 ,
q0

and Dc = 1.6996 × 10−5 . We assume that control values are saturated at u0 = p0 .

Then, applying the stability results presented in Theorem 3.2 and using the algorithm proposed
in (3.26). For 0.1 ≤ d(k) ≤ 0.5, μ = 1, and T = 0.1, the corresponding matrices of (2.7) and
prescribed scalar γ are Aa = 0.2, Ba = 0.03, Ca = 0.05, Da = 0.1, and γ = 0.01, respectively.
The performance in reference queue tracking and disturbance rejection has been investigated
through simulations. Two well-known AQM methods, RED and REM, were also simulated for
comparison, with the parameters listed in Table 2. The transfer function for RED (Floyd and
KRED
Jacobson, 1993) is CRED (s) = LRED s+K . For REM, the end-to-end marketing probability
 RED
is p (loge ωREM ) ρ(kT ) where the update of the price ρ(kT ) in period T follows equation

49
 International Journal of Ecological Economics & Statistics

(Athuraliya et al., 2001)

   
ρ((k + 1)T ) = max ρ(kT ) + η b((k + 1)T ) − 0.99b(kT ) − 1.75 , 0

Table 2: Controllers parameters.

Controllers Parameters value
RED LRED = 1.86 × 10−4 , KRED = 0.005
REM ωREM = 1.001, η = 0.001

Thus, using the algorithm proposed in (3.26) and the initial values ξ0 = [10 − 10]T , the queue
size regulation and deviations of drop probability are shown in Figure 3, when the Gaussian
noise in Figure 2 is used as disturbance to check their effect.

0.08

0.06

0.04

0.02

-0.02

-0.04

-0.06

-0.08

-0.1
0 5 10 15
Time (s)

Figure 2: Random disturbance.

The simulation results show the performance of the proposed controller: it manages to main-
tain the queue length at the target value despite the inherent time-varying dynamics. On the
contrary using a standard RED controller the queue length is far from the desired value, pre-
senting a sluggish response, resulting in degraded utilization, losses, and high variance of
queuing delay. Moreover, REM presents high variations, which in turn results in high and fluc-
tuant values of the RTT, affecting the performance of the network and aggravating the quality
of services. The drop rate (given by the rate probability) is smaller with the proposed controller
than with standard RED and REM.

5 Conclusion

An AQM approach for congestion control has been discussed, that incorporates explicitly anti-
windup to incorporate the effect of saturation of the control signal. The approach is based on a
discrete-time TCP flow model with link capacity disturbance and time-varying delay: a sufficient
condition for stability is proposed, which is then incorporated into an optimization algorithm to
design anti-windup that ensure stability for the largest set of admissible initial states. Some
simulation results are presented to demonstrate the effectiveness of the proposed approach.

50
 International Journal of Ecological Economics & Statistics

Figure 3: Variation over average value of queue (left) and discard probability (right).

Acknowledgment

We would like to thank Prof. Tadeo for many helpful discussions and Prof. Bender for providing
assistance with the software. The last author has been funded by DPI2014-54530-R.

References

Athuraliya, S., Low, S. H., Li, V. H. and Yin, Q. 2001 . REM: active queue management, IEEE
Network 15(3): 48–53.

Bender, F. A. 2013 . Delay dependent antiwindup synthesis for time delay systems, Int. J.
Intelligent Control Systems 18(1): 1–9.

Chaibi, N., Tissir, E. H., Hmamed, A. and Idrissi, S. 2013 . H∞ performance to stability and sta-
bilization for singular time delay systems, Int. J. Ecological Economics Statistics 28(1): 18–
29.

51
 International Journal of Ecological Economics & Statistics

Chen, C.-K., Hung, Y.-C., Liao, T.-L. and Yan, J.-J. 2007 . Design of robust active queue
management controllers for a class of TCP communication networks, Information Sci.
177(19): 4059–4071.

El Fezazi, N., El Haoussi, F., Tissir, E. H. and Tadeo, F. 2015. Delay dependent anti-windup
synthesis for time-varying delay systems with saturating actuators, Int. J. Computer Appl.
111(1): 1–6.

El Fezazi, N., El Haoussi, F., Tissir, E. H., Husain, A. R. and Zakaria, M. I. 2016. Delay-
dependent robust anti-windup synthesis approach to AQM in TCP/IP networks, 2nd IEEE
Int. Conference on Electrical and Information Technologies, Tanger, Morocco 294–299.

El Fezazi, N., Lamrabet, O., El Haoussi, F., Tissir, E. H., Alvarez, T. and Tadeo, F. 2016. Robust
controller design for congestion control in TCP/IP routers, 5th IEEE Int. Conference on
Systems and Control, Marrakesh, Morocco 243–249.

El Fezazi, N., El Haoussi, F., Tissir, E. H., Alvarez, T. and Tadeo, F. 2017. Robust stabilization
using LMI techniques of neutral time-delay systems subject to input saturation, In J. of
Physics: Conference Series, IOP Publishing, Lille, France 783(1): 012031.

El Fezazi, N., Tissir, E. H., El Haoussi, F., Alvarez, T. and Tadeo, F. 2017. Control based
on saturated time-delay systems theory of Mach number in wind tunnels, Circuits, Syst.,
Signal Processing 1–18.

El Haoussi, F. and Tissir, E. H. 2007 . Delay-dependent robust observer and H∞ based con-
troller design for neutral systems, Int. J. Ecological Economics Statistics 9(F07): 57–64.

El Haoussi F., Tissir E. H. and Tadeo, F. 2014 . Advances in the robust stabilization of neutral
systems with saturating actuators, Int. J. Ecology Development 28(2): 49–62.

Floyd, S. and Jacobson, V. 1993 . Random early detection gateways for congestion avoidance,
IEEE/ACM Trans. Networking 1(4): 397–413.

Hollot, C.V., Misra, V., Towsley, D. and Gong, W.-B. 2001 . On designing improved controllers
for AQM routers supporting TCP flows, 20th IEEE Int. Conference on INFOCOM, Alaska,
USA 3: 1726–1734.

Hollot, C.V., Misra, V., Towsley, D. and Gong, W.-B. 2002 . Analysis and design of controllers
for AQM routers supporting TCP flows, Trans. Automatic Control 47(6): 945–959.

Manfredi, S., Di Bernardo, M. and Garofalo F. 2009. Design, validation and experimental testing
of a robust AQM control, Control Engineering Practice 17(3): 394–407.

Misra, V., Gong, W.-B. and Towsley, D. 2000 . Fluid-based analysis of a network of AQM routers
supporting TCP flows with an application to RED, IEEE Int. Conference on SIGCOMM,
Stockholm, Sweden 30(4): 151–160.

Negi, R., Purwar, S. and Kar, H. 2012 . Delay-dependent stability analysis of discrete time
delay systems with actuator saturation, Intelligent Control Automation 3(1): 34.

52
 International Journal of Ecological Economics & Statistics

Ott, T. J., Lakshman, T. V. and Wong, L. H. 1999 . Sred: stabilized RED, 18th IEEE Int.
Conference on INFOCOM, New York, USA 3: 1346–1355.

Sun, J., Chen, G., Ko, K. T., Chan, S. and Zukerman, M. 2003 . PD-controller: a new active
queue management scheme, IEEE Int. Conference on Global Telecommunications, San
Francisco, USA 6: 3103–3107.

Tarbouriech, S., Da Silva, J. M. G. and Garcia, G. 2004 . Delay-dependent anti-windup strategy

for linear systems with saturating inputs and delayed outputs, Int. J. Robust Nonlinear
Control 14(7): 665–682.

Tissir, E. H. 2009 . On Delay Dependent Stability Robustness for Systems With Uncertain
Time Varying Delays, Int. J. Ecological Economics Statistics 14(S09): 87–97.

Tissir, E. H. 2014 . Anti-windup design for singular systems with saturating actuators, Int. J.
Ecological Economics Statistics 32(1): 66–74.

Wang, C., Li, B., Hou, Y. T., Sohraby, K. and Lin, Y. 2004 . LRED: a robust active queue man-
agement scheme based on packet loss ratio, 23th IEEE Int. Conference on INFOCOM,
Hong Kong 1.

Yanfie, F., Fengyuan, R. and Chuang, L. 2003 . Design a PID controller for active queue
management, 8th IEEE Int. Conference on Computers and Communications, Kemer,
Turkey 985–990.

Yan, P., Gao, Y. and Ozbay, H. 2005 . A variable structure control approach to active queue
management for TCP with ECN, Trans. Control Systems Technology 13(2): 203–215.

Zhang, P., Ye, C. Q., Ma, X. Y., Chen, Y. H. and Li, X. 2007 . Using Lyapunov function to design
optimal controller for AQM routers, J. Zhejiang University-Science A 8(1): 113–118.

53

View publication stats