ISA Transactions 47 (2008) 143–155

www.elsevier.com/locate/isatrans

AQM controller design for networks supporting TCP vegas: A control

theoretical approach
Nooshin Bigdeli, Mohammad Haeri ∗
Advance Control System Lab., Electrical Engineering Department, Sharif University of Technology, Tehran, Iran

Received 21 February 2007; received in revised form 14 May 2007; accepted 31 May 2007
Available online 30 July 2007

Abstract

In this paper, a mathematical model and control theoretical framework for designing AQM controllers in networks supporting TCP Vegas
is introduced. We have emphasized on a modified TCP Vegas algorithm that can respond to congestion signals through explicit congestion
notification (ECN). The overall nonlinear delayed differential equations of the dynamics model of closed loop system have been derived based
on TCP Vegas model. The model is then linearized to derive a transfer function representation between the packet marking probability and the
bottleneck router queue length as the input and output of the modified TCP Vegas/AQM system. The model properties have been then examined
especially for the case of single bottleneck homogeneous network which is closely investigated. Finally an AQM controller based on Coefficient
Diagram Method (CDM) has been designed for the system and its performance has been compared with some other AQM controllers. CDM
is a new indirect pole placement method that considers the speed, stability and robustness of the closed loop system in terms of time domain
specifications. In order for synthesizing the simulation scenarios, our campus router traffic has been studied experimentally for a sample period of
one hour and the corresponding parameters has been extracted. The simulation results are representative of good performance of developed TCP
Vegas/AQM structure for different simulated scenarios.
c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

Keywords: Active queue management; Coefficient diagram method; Computer networks; Congestion control; Internet measurement; Network modelling; TCP Vegas

1. Introduction Almost all of the responsive flows in current Internet

In current Internet, unmanaged congestion results in vast
lost of resources and data that may cause serious problems
in different levels of Internet usages such as economy and
trading, aerology, transport, etc. Therefore, employing some
algorithms to avoid or manage this phenomenon in order to
guarantee some quality of service (QoS) requirements seems
inevitable. Technically speaking, there are two parts of network
that may deal with the congestion control problem. In end-
users, responsive flows adjust their data sending rate according
to the network congestion status; and, in the routers some
algorithms are employed to send some congestion notification
signals to the users. These congestion notification signals are
based on the congestion status, available resources, and the QoS
requirements in the corresponding router.
∗ Corresponding author. Tel.: +98 21 66165964; fax: +98 21 66023261.
E-mail addresses: bigdeli@ee.sharif.edu (N. Bigdeli), haeri@sharif.edu
(M. Haeri).
use Transmission Control Protocol (TCP) or TCP-friendly
protocols as their congestion control algorithm. TCP, by
regulating the data sending rate in different congestion
control phases (slow start, congestion avoidance, and fast
recovery/retransmit), tries to prevent or reduce the congestion in
the Internet [1]. So far, lots of TCP variants have been proposed
to enhance the congestion control capability.
One of the TCP variants, the TCP Vegas, was first introduced
by Brakmo and Peterson [2] in 1995 and then improved by
Bonald [3] and Mo et al. [4] as an alternative to commonly used
TCP Reno. In TCP Reno algorithm, the congestion measure
is induced through marking or dropping packets in congested
routers. In contrast, a Vegas source anticipates the congestion
by monitoring the difference between the rate it is expecting
to see and the rate it is actually realizing. Vegas’ strategy
is to adjust the source’s sending rate in an attempt to keep
a small number of packets buffered in the routers along the
path. Vegas does not support any reaction to explicit congestion
notification signals from the network and only reacts to implicit

c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

0019-0578/$ - see front matter

doi:10.1016/j.isatra.2007.05.001
144 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155
feedback from the network. This property, however, leads to
some types of performance degradation in cases that the round
trip propagation delay can not be calculated accurately. Some
examples are when there is backlogged data in the queue or
in case of rerouting [5,6]. There are some improved Vegas
algorithms proposed in the literature that attempt to treat TCP
Vegas drawbacks [7,8]. In [7], a modified Vegas called Vegas-
Like AMP-TCP has been introduced which employs model
predictive control strategy in the end-hosts to regulate their
number of in-flight packets such that each flow derives its
desired round trip time value. Another proposal called Vegas-
A has been introduced in [8] to address the limitations of TCP
Vegas such as its problem in case of rerouting. In this method,
the desired number of in-flight packets is adaptively determined
based on actual throughput of TCP for each flow. However,
none of these methods use explicit/implicit feedback signals
from congestion status of the underlying networks, while it has
been shown that by cooperation of end-users and intermediate
nodes (routers) in the network, the overall behavior of the
networks can be noticeably improved and congestion can be
prevented or treated more adequately [6].
Active Queue Management (AQM) methods are those
techniques which attempt to prevent the congestion and
regulate the queue length by sending congestion signals through
marking/dropping packets, which is assumed to eventually
cause the senders reduce their sending rates. They try to
solve the congestion control problem by regulating their queue
length around a desired value [9,10]. There are different AQM
methods introduced in the literature. They are designed such
that they can cooperate with the TCP variants that can react
to congestion notification signals [6]. Many of these AQM
methods are designed based on control theory methods and use
the fluid flow models of different TCP variants especially for
the TCP Reno model of [11]. For example in [12], proportional
plus integral (PI) controller and in [13] an adaptive stable PI
has been designed based on this model. Another example is the
coefficient diagram method controller designed by the authors
in [14–16]. In the same way, by modifying Vegas algorithm so
that it reacts to the congestion signals coming from the routers,
one can expect more precise reaction as well. In [6], a modified
Vegas is proposed, in which the information of packet marking
probability in the queues in addition to Round Trip Time (RTT)
is used to estimate the propagation round-trip delay and to
determine the data sending rate, when random early marking
(REM) algorithm [9] is employed in routers as AQM method.
In this paper, the proposed idea in [6] has been extended
to derive a framework for designing AQM methods for network
supporting TCP Vegas. We have emphasized on a modified TCP
Vegas algorithm that can respond to congestion signals through
explicit congestion notification (ECN). If networks support
the ECN, AQM schemes mark the ECN bits as congestion
notification [17] instead of dropping packets. We first formulate
the model of the closed loop TCP Vegas/AQM network in its
general form for an arbitrary number of bottlenecks. Then this
model is linearized to obtain a transfer function representation
of the small signal behavior of TCP Vegas/AQM networks.
The transfer functions relate small variations in the packet
marking probabilities and queue lengths in bottleneck routers
as system inputs and outputs, respectively. The derived model
is then used to analyze system behaviour and to design AQM
controller. In special case of a single bottleneck shared by
some homogeneous flows, the developed model is similar to
one derived in [9] for TCP Reno/AQM networks by Misra
et al. As part of our work, we use this model to investigate
the homogeneous single bottleneck system properties more
closely in terms of topology and traffic parameters. Afterwards,
design of an AQM controller is performed based on Coefficient
Diagram Method (CDM). The design procedure of CDM as
an AQM controller for TCP Reno/AQM networks have been
discussed by the authors in series of papers [14–16] and its
out-performance with respect to common AQM controllers
has been shown there. This good background motivated the
authors to design the CDM as AQM controller for networks
supporting TCP Vegas. In this paper, the performance of
the TCP Vegas/CDM networks is evaluated through different
simulation scenarios in comparison with other AQM methods
such as RED and PI. In order for synthesizing our simulation
scenario, the properties of the traffic passing through our
campus router have been measured via some experiments.
Packet length and RTT, traffic composition, and flow size
and number of active flows are the considered features of the
measured data. The data have been gathered for one hour on
November 20 2006, 9–10 AM.
The remainder of this paper is organized as follows. In
Section 2, the principals of TCP Vegas operation is described.
The modified Vegas and its relationship with AQM control
signal, the linearized model and special case of homogenous
networks are presented in Section 3. Section 4 deals with CDM
design as an AQM controller. Simulation results are brought in
Section 5. The paper is concluded in Section 6.
2. TCP vegas principals
In this section, the principals of the operation of a general
network supporting TCP Vegas are presented [5]. Consider a
network consisting of a set of L links with finite capacities
cl , l ∈ L. Assume the links are shared by a set of N sources
indexed by r . Each source r attains a utility Ur (xr ) when
it transmits its data packets with a rate of xr (in packets/s)
through a set of links defined by the L × N routing matrix

1 if source r uses link l
Rlr =
0 otherwise.
Associated with each link l is a congestion measure pl (t),
called ‘price’, which is the scaled queuing delay at link l. Let us
denote the total round trip propagation delay of source r by dr ,
the forward delay from source r to link l by τrl , the backward
delay from link l to source r by τlr , and the round trip time by
Tr . Then at time t, the aggregate price in the path of source r
will be:
X
qr (t) := Rlr pl (t − τlr ), (1)
l
 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155 145

and the link l observes the aggregate source rate of 3.1. Modified vegas/AQM networks modelling
X
yl (t) := Rlr xr (t − τrl ), (2) Consider the modified Vegas in [6], in which the variation in
r window size is a function of dropping/marking probability of
packets in the intermediate routers, that is
where, Tr = τrl + τlr , ∀l ∈ Lis the round trip time.
According to an interpretation of Vegas, a source monitors 1 qr (t) + εr

 wr (t) + if xr (t) < αr
the difference between its expected rate and its actual rate, and (t)

T dr


 r
increases or decreases its window size by one in the next RTT

wr (t + 1) = 1 qr (t) + εr (6)
according to weather the difference is less than or greater than wr (t) − if xr (t) > αr
T (t) d

r r
a parameter αr [5]. Here, αr is a Vegas protocol parameter in




wr (t) otherwise,

terms of KB/s and has significant effect on the fairness of Vegas
protocol [5]. Let wr (t) be the window size of source r at time t, where εr is a function of the aggregate marking/dropping
then the change in window size can be modeled by one packet probability, m r , along the dedicated path for source r , that is
per RTT according to
εr = µr m r , (7)
1 wr (t) wr (t)

 wr (t) + if − < αr where µr is a constant that should be chosen. m r can be
(t)



 Tr dr Tr approximated as
wr (t + 1) = 1 wr (t) wr (t) (3a)
w (t) > α
Y


 r −
Tr (t)
if
dr
−
Tr
r m r (t) = 1 − Rlr (1 − pd,l (t − τlr ))

wr (t) else . l

X
∼
= Rlr pd,l (t − τlr ), (8)
In [2], wr (t)/dr is referred to as the expected rate, l
wr (t)/Tr (t) as the actual rate, and the difference wr (t)/dr −
w where pd,l is the marking/dropping probability for the queue in
Pr (t)/Tr (t) = xr (t)q Pr (t)/dr as ‘DIFF’; where qr (t) = link l. By employing such algorithm, the TCP Vegas reacts to
l Rlr pl (t − τlr ) = l Rlr bl (t − τlr )/cl is the end-to-end
queuing delay (backlog) in path of source r and bl (t) is the the congestion notification signals fed back by the router AQM
queue length of link l [5]. But, the dynamics of queue length controllers.
As our complementary point, we argue that many effective
variation in link l can be considered as the difference between
AQM methods such as RED [10] use the exponentially
its input data rate and the rate of servicing data. If the data
weighted moving average (EWMA) of the queue length.
processing is not the bottleneck in the corresponding router, the
In [18], the necessity and usefulness of such assumption is
queue length dynamic can be modelled as
discussed in details. The average queue length bavg,l (t) for
ḃl (t) = yl (t) − cl . (3b) the queue l as an exponentially weighted moving average of
instantaneous queue length bl (t) with a weight factor ρl is [10]:
From (3) and replacing the sending rate xr (t) = Z t
wr (t)/Tr (t), the Vegas algorithm can be described by the bavg,l = ρl es−t bl (s)ds. (9)
following couple of differential equations: −∞
Replacing (6) into (5), and from (3b) and (9), the overall
(yl (t) − cl )/cl if pl (t) > 0

ṗl (t) = (4) nonlinear model of the interaction between TCP Vegas sources
(yl (t) − cl )+ /cl if pl (t) = 0, and queue length dynamic in intermediate routers can be
ẋr (t) = sgn(1 − xr (t)qr (t)/αr dr )/Tr2 (t), (5) derived as three sets of nonlinear differential equations as
qr (t) + µr m r (t)
 
where, (z)+ = max(0, z) and sgn(z) is the sign function [5]. ẋr = sgn 1 − xr (t) /Tr2 (t)
αr dr
(10)
3. Modified vegas and AQM design framework ḃl (t) = (yl (t) − cl )+
ḃavg,l (t) = −ρl bavg,l (t) + ρl bl (t),
The performance of Vegas algorithm would be improved if
where pl (t), the average queuing delay, is now defined as
a source can make use of marking information of the routers
(those employ AQM) to estimate the queue occupancy of pl (t) = bavg,l (t)/cl , (11)
the bottleneck [6]. On the other hand, by employing AQM b (t−τ )
and Tr (t) = qr (t) + dr = l Rlr avg,l cl rl + dr . The other
P
methods in routers, one can regulate the queuing delay and
therefore, the QoS requirements is satisfied by controlling definitions remain unchanged.
factors such as average queuing delay and its jitter or supporting
3.2. Model linearization and the transfer function representa-
differentiated services. In [6], the first aspect, i.e. improving
tion
Vegas performance has been dealt with. Here, we use the same
context to achieve the second goal, i.e. deriving a framework for In this section, the derived model is analysed in order to
designing AQM methods in order to satisfy QoS requirements. design the AQM congestion controller. To do so, we first derive
146 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155

the equilibrium point of the system from an AQM design point

of view. If the desired queue length in link l is assumed to be
Bl∗ , then the network parameters in the equilibrium satisfy the
following relationships:
αr dr X B∗ X
xr∗ = ∗ , qr∗ = Rlr l , yl∗ = Rlr xr∗ ,
qr + µt m r
∗
l
cl r
∗
bavg,l = Bl∗ . (12)
If we approximate the sign function as in [19]
Fig. 1. Single bottleneck topology.
sgn(z) ≈ 2 tan−1 (ηz)/π,
where the approximation becomes exact in the limit as η → ∞,
the first relation in (10) can then be approximated by where the vectors δy, δx, δq, δbavg , δm, and δpd consist of
elements δyl , δxr , δqr , δbavg,l , δm r , and δpd,l respectively, and,
ẋr (t) = fr (xr (t), qr (t), m r (t))  −sτ
e rl if Rlr = 1
qr (t) + µr m r (t) Rlr (s) =
 
2 1
= tan −1
η 1 − x r (t) . (13) 0 otherwise.
π Tr2 (t) αr dr
Now, let’s assume Substituting (14c) into (14d) we derive

xr (t) = xr∗ + δxr (t), qr (t) = qr∗ + δqr (t), ρl

 
δbavg (s) = diag R(s)δx(s). (14g)
bl (t) = Bl∗ + δbl (t), bavg,l (t) = bavg,l
∗
+ δbavg,l (t) s(s + ρl )
m r (t) = m r∗ + δm r (t), pd,l (t) = pd,l
∗
+ δpd,l (t), Replacing (14e)–(14g) into the vector form equivalent of (14b),
yl (t) = yl∗ + ∂ yl (t). the open loop linearized model for the control design purpose
is derived as
Then, the linearized model of the networks at the sources can
ρl
  
be expressed as s I + diag R(s)
cl (s + ρl )
∂ fr  ∂ fr  ∂ fr 
  
ẋr = δ ẋr = δx δq δm r , βr ar e−sTr
  
r + r + (14a)
∂ x r ∗ ∂qr ∗ ∂m r ∗ × diag R T
(−s) δbavg
sTr∗ + ar
where,  
ρl

∂ fr 

∂ fr 
 = − diag R(s)
=−
2η 1
, =−
2η 1
, (s + ρl )
∂ xr ∗ π xr Tr
∗ ∗2 ∂qr ∗  π (qr + µr m r∗ )Tr∗2
∗
µr βr ar e−sTr
  
∂ fr  µr × diag R (−s) δpd ,
T
(15)

2η
=− and, Tr∗ = dr + qr∗ . sTr∗ + ar
∂m r ∗  π (qr + µr m r∗ )Tr∗2
∗
x∗
In the frequency (Laplace) domain we have where, βr = q ∗ +µr r m ∗ and I is the identity matrix of proper
r r
dimension. Note that the minus sign in the above equation is
xr∗ ar representative of the reverse effect of drop probability on the
δxr (s) = − (δqr (s) + µr δm r (s)), (14b)
qr∗ + µr m r∗ Tr∗ s + ar queue length. By defining
where ar = 2η/π Tr∗ . At the links, the linearized equations
δbl (t) =: Bl∗ − bl (t), (16)
become
δ ḃl = δyl , the minus sign is omitted. On the other hand, to derive (15) only
the congested links are involved, therefore, the dimension of the
δ ḃavg,l = −ρl δbavg,l + ρl δbl .
model is generally smaller than L.
Therefore, we have:
ρl 3.3. Homogeneous networks model and its properties
δbavg,l (s) = δyl (s). (14c)
s(s + ρl )
Eqs. (14a) and (14b) can now be written in the matrix form
In this section, we apply the results of the previous sections
with the use of definitions given in (1), (2) and (11). We have
to investigate the properties of the resulted model for a network
δy(s) = R(s)δx(s), (14d) consisting of N p homogenous long-lived FTP sources that
 −sTr 
e support the modified Vegas described in the last section and
δq(s) = diag R T (−s)δbavg (s), (14e) share a single bottleneck of capacity C p as in Fig. 1. It is
cl
assumed that all the flows experience the same round trip
δm = diag{e−sTr }R T (−s)δpd , (14f) propagation delay d p . For such a system, the representing
 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155 147

Fig. 3. Block diagram of the small signal interactions in the linearized model
of Eq. (19).

Fig. 2. Block diagram of the interactions among different parameters in the

nonlinear model of Eq. (17).

model equations become

b̄ p
 
1 C p + µ p pd (t − τb )
ẋ p (t) = 2 sgn 1 − x p (t) 
T p (t) αpdp
(17) Fig. 4. The function sgn(z) versus the function 2π −1 tan−1 (ηz) for different
ḃ p (t) = (N p x p (t − τ f ) − C p )+ values of η.

b̄˙ (t) = −ρ b̄ (t) + ρ b (t),

p p p p p
Neglecting the queue averaging procedure, the transfer
where, x p , b p , b̄ p , τ f , and τb stand for TCP sending rate function of (20a) will be
for each source, bottleneck queue length, bottleneck average ∗
queue length, the forward and backward delay, respectively. δb p C p N p µ p β p a p e−T p s
G r (s) = =− ∗ .
T p is the RTT and we have: T p = d p + b̄ p /C p . The other δpd (s) T p∗ C p s 2 + a p C p s + N p β p a p e−T p s
parameters in (17) such as µ p , α p , and ρ p represent the (20b)
network parameters as defined before. The block diagram of
Fig. 2 shows the interactions among different parameters in the Considering the above relationships and the block diagram
differential equations in (17). In this figure, the TCP Vegas rate of Fig. 3 one can conclude
adjustment and the queue blow up dynamics are highlighted 1. TCP Vegas algorithm by itself serves as a delayed system
clearly. The equilibrium point of the system for a desired queue with a proportional feedback controller of gain 1/C p .
length b∗p = B0 will be as Removing this feedback from the system, the original open
Cp αpdp N p B0 loop system has two poles at s = 0 and s = −a p /T p∗ , and
b̄∗p = B0 , x ∗p = , pd∗ = − . (18) one pole at s = −ρ p due to the queue averaging component.
Np µpC p µpC p
Therefore, once the averaging weight is selected, the AQM
Then, the small signal behavior of such system can be controller can be designed to reach the desired queue
represented in Laplace domain as regulation properties.
∗ ! 2. Before designing the controller, one should decide about the
x ∗p a p e−T p s δ b̄ p model parameters η, µ p , and ρ p . To choose a proper value
δx p = − + µ p δpd
B0 + µ p pd∗ T p∗ s + a p C p for η, Fig. 4 where the sgn(z) and 2π −1 tan−1 (ηz) have been
∗ ! plotted for different values of η (10–100 with step 10) is
β p a p e−T p s δ b̄ p considered. For η ≥ 50, difference between 2π −1 tan−1 (ηz)
= − ∗ + µ p δpd (19)
Tp s + a p Cp and sgn(z) is negligible.
δ ḃ p = N p δx p Parameter µ p can be determined from equilibrium
point requirements in a procedure like follows. From the
δ b̄˙ p = −ρ p δ b̄ p + ρ p δb p . equilibrium point given in (18), once B0 is selected, b̄∗p and
The block diagram of Fig. 3 shows the small signal x ∗p will be uniquely determined (for the nominal value of
interactions of system in (19). From this figure and from (15), N p ). pd∗ depends on µ p in addition to the network and TCP
the small signal transfer function of the system becomes Vegas parameters. Therefore, µ p is chosen such that pd∗ has
a reasonable value for the nominal or the worst case of the
δ b̄ p network parameter values.
G(s) =
δpd (s) 3. The last parameter to be chosen is ρ p , which is the pole of
∗
C p N p µ p ρ p β p a p e−T p s the queue averaging transfer function
=− ∗ .
T p∗ C p s 3 + (a p + T p∗ ρ p )C p s 2 + ρ p a p C p s + N p ρ p β p a p e−T p s b̄ p (s) ρp
= . (21)
(20a) b p (s) s + ρp
148 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155
Fig. 5. AQM as feedback control.
This transfer function represents a low-pass filter with steady
state gain of 1. Application of the filter results in the
bandwidth reduction which means reduction in the queue
fluctuations and therefore improving the global stability [6].
The smaller ρ p causes the less fluctuating and the more
sluggish system. Therefore, choice of ρ p is a trade-off
between these two phenomena. On the other hand, transfer
function of (20a) can be rewritten as (with use of (16))
∗ 4. Transient response: In last section we mentioned that TCP
µ p N p ρ p a p β p e−T p s
(T p∗ s+a p )(s+ρ p )
G(s) = ∗ .
N p ρ p a p β p e−T p s
s+ C p (T p∗ s+a p )(s+ρ p )
In order to design the AQM controller, the effect of
the transport delay can be neglected if the term (T p∗ s +
a p )(s + ρ p ) has much lower bandwidth than the delay term
bandwidth, 1/T p∗ . This can be done by choosing ρ p  to be implemented in high-speed networks.
1/T p∗ . Therefore, the linearized model in (20a) is reduced
to
G(s)
µp N pρpapβpC p
=
T p∗ C p s 3 + (a p + T p∗ ρ p )C p s 2 + ρ p a p C p s + N p ρ p a p β p
n0
= .
d3 s + d2 s 2 + d1 s + d0
3
Based on this discussion, the block diagram of AQM as
feedback controller for system of (23) has been shown in Fig. 5,
where the delay effect is neglected due to the proper choice of
ρp.
4. AQM controller design
In this section, the AQM controller design procedure is dealt
with. The main goals of AQM design are [12]:
1. Efficient queue utilization: For efficient use, the queue
should avoid overflow or emptiness. The former situation
results in packets loss and undesired retransmissions while
an empty buffer under-utilizes the link. Therefore, both of
these extremes should be avoided both in transient and
steady state operation.
2. Queuing delay and jitter: The time required for a data packet
to be serviced by the routing queue is the queuing delay.
This time, together with the propagation delay, accounts for
the network delay. It is desirable to keep both the queuing
Fig. 6. The block diagram of CDM control system.
delay and its variations (jitter) small. This calls for regulating
the queue length around small values. However, doing so
may result in link underutilization. Therefore, there is a
fundamental trade-off in AQM design.
3. Robustness: AQM schemes need to maintain closed loop
performance in face of varying network conditions. These
conditions include variations in both the topology and traffic
properties of the network. Topology parameters are the link
propagation delay and bottleneck link capacity and the main
traffic property is the number of TCP sessions.
Vegas has no control on the transient response properties
(22) of the closed loop system. The response may be sluggish
or showing overshoot in transients. Therefore, the control-
theory-based AQM controllers are proposed to consider such
properties as well.
5. Simplicity: The designed controller should be simple enough
To meet these specifications, the Coefficient Diagram
Method (CDM) will be used to design AQM controller for TCP
Vegas/AQM networks in this paper. In continue details of the
design procedure are discussed.
4.1. Coefficient diagram method
(23)
The block diagram of CDM as a control structure is depicted
in Fig. 6. As observed, the CDM is an algebraic controller
design approach that is expressed in polynomial form. In
CDM the closed loop characteristic and controller polynomials
are designed simultaneously such that some desired closed
loop performance requirements are satisfied [20–24]. The
performance issues are the speed of response, stability and
robustness which are represented in terms of the closed loop
characteristic polynomial coefficients via some indices. The
basic idea of CDM was developed in the 1960s by Naslin and
in the context of Characteristic Ratio (CR) and characteristic
pulstance [22]. It has been shown that an appropriate behaviour
of a system can be expected by choosing certain ratios of
coefficients of the characteristic polynomial. The basis of CDM
has been known in industry since 1970 and some successful
applications have been reported in servo control, steel mill
drive control, gas turbine control, and spacecraft control [22].
The today’s form of the CDM was first suggested by Manabe
in 1994 [21] by conjunction of Naslin idea and Lipatov and
Sokhlov stability criteria [25] and then completed by the same
researcher in 1998 [20]. CDM design is based on the stability
 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155 149

indices, γi , the stability limits, γi∗ , and the equivalent time

constant, τ , and for a specified settling time, a controller of
the lowest order with the widest bandwidth can be easily
designed.

4.1.1. Mathematical relationships

Consider the feedback system shown in Fig. 6 [23]. Here
ys is the output signal, u s is the controller output, rs is the
reference input, ds represents the disturbance and/or uncertainty
in the plant model, and n s denotes the measurement noise. The
numerator and denominator of the plant transfer function are
given by N (s) and D(s) respectively. B(s) and A(s) are the
numerator and denominator of the controller transfer function.
F(s) is called the reference numerator. The output of the Fig. 7. A coefficient diagram.
controlled system is
N (s)F(s) A(s)N (s) N (s)B(s) line connecting 1 to τ . Diagram in Fig. 7 indicates coefficient
ys = rs + ds − ns , (24)
P(s) P(s) P(s) diagram for the characteristic polynomial: P(s) = 0.5s 5 +
where P(s) is the characteristic polynomial and is given as 1.8s 4 + 2.5s 3 + 2.5s 2 + 1.5s + 0.4. The CDM parameters
are γi = [∞, 2.5, 2, 2, 2, ∞], γi∗ = [∗, 0.5, 1, 0.9, 0.5, ∗],
n
X and τ = p1 / p0 = 3.75 in this case. By increasing the
P(s) = D(s)A(s) + N (s)B(s) = pi s i . (25) curvature of pi ’s, the closed loop system becomes more stable.
i=0
This corresponds to the larger stability index γi . When this
For this characteristic polynomial, the stability index γi , curve is right side down, the equivalent time constant τ that
the equivalent time constant, τ , and the stability limit γi∗ , are represents the speed of the closed loop system response is
defined as small and therefore the closed loop system has fast dynamics
[21].
pi2
γi = , i = 1, . . . , n − 1, γ0 = γn = ∞
pi+1 pi−1 (26) 4.1.3. Stability and robustness
τ = p1 / p0 , γi∗ = γi−1
−1
+ γi+1
−1
. The stability and robustness conditions of the control
For a set of given stability indices and the equivalent system based on CDM are discussed in series of the research
time constant, the characteristic polynomial parameters are papers [20,22,26,27]. Based on the Routh–Hurwitz stability
determined as criterion, the stability condition for a third order system is

pi+1 pj 1 p2 p1 > p0 p3 → γ1 γ2 > 1. (29a)

= , i> j
pi p j−1 γi γi−1 · · · γ j+1 γ j
(27) For a fourth order system, this criterion results in the
p0 τ i
pi = . following condition:
γi−1 γi−2
2 ···γ
1
i−1
p1 p4 p0 p3
Using these relations one can write the characteristic p2 > + → γ2 > γ2∗ . (29b)
p3 p1
polynomial of the system in terms of the design parameters p0 ,
τ , and γi : Stability and instability conditions for the 5th and higher
     order systems are discussed in [23] and are given in (30a) and
X n i−1
Y 1  (30b) respectively
P(s) = p0    (τ s) + τ s + 1 .
i
(28)
 i=2 j=1 γ j
γi > 1.12γi∗ i = 2, . . . , (n − 2) (30a)

i− j
γi γi+1 ≤ 1 i = 2, · · · , (n − 2). (30b)
4.1.2. Coefficient diagram The robustness, as a measure of how fast the poles of
The heart of the CDM controller is the semilogarithmic the closed loop system move toward the imaginary axis
coefficient diagram depicted in Fig. 7, where three performance because of the system parameters variation, can be achieved as
specifications namely the stability index, γi , the equivalent time generally by making some sacrifice of the closed loop system
constant, τ , and the stability limit γi∗ are depicted versus degree performance. Therefore, relaxed robustness condition is given
of polynomial coefficient i. The performance specifications as
γi , τ , and γi∗ are read by the right side scale where the
coefficients of the characteristic polynomial pi , is read from γi > 1.5γi∗ . (30c)
the left side scale. The time constant, τ , is expressed by a
150 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155

Table 1 the controller design procedure described in the previous

Degree requirements for CDM control polynomials section, the proper controller polynomials A(s) and B(s) is
Type of input disturbance of order three and therefore, the closed loop characteristic
None Step Ramp Impulse/sinusoidal polynomial is of order six. In this case we will have one
Deg(A) nd − 1 nd nd + 1 nd − 1 redundant parameter or equivalently one degree of freedom in
Deg(B) nd − 1 nd nd + 1 nd − 1 designing the CDM controller. For better disturbance rejection
Condition – a0 = 0 a0 = 0, a1 = 0 – capability, we choose a0 = 0 in A(s) polynomial. Therefore:
Deg(P) 2n d − 1 2n d 2n d + 1 nd − 1
A(s) = a3 s 3 + a2 s 2 + a1 s
(34)
4.1.4. CDM controller design B(s) = b3 s 3 + b2 s 2 + b1 s + b0 .
For an arbitrary system, polynomials N (s), D(s), A(s), B(s)
and F(s) (Fig. 7) are considered as follows: Choosing a proper equivalent time constant τ is also a
challenge. Because of existing delay in the network dynamics,
N (s) = n n n s n n + · · · + n 1 s + n 0 , the time specifications of the closed loop system can not be
D(s) = dn d s n d + · · · + d1 s + d0 chosen arbitrarily small. Therefore, the time constant should
also be chosen large enough. Here, time constant about 4T p∗
A(s) = an a s n a + · · · + a1 s + a0 , (31) is recommended. Stability indices should also be chosen large
nb
B(s) = bn b s + · · · + b1 s + b0 enough to handle time varying properties of the network
F(s) = f n f s nf
+ · · · + f1s + f0. parameters.
The closed loop characteristic equation P(s) is then given as
5. Simulations
n
X
P(s) = D(s)A(s) + N (s)B(s) = pi s i 5.1. The simulation scenario
i=0

= (dn d an a )s n d +n a + · · · + (d0 a0 + n 0 b0 ). (32) The performance of the CDM as an AQM controller

is evaluated through simulations using MATLAB. We have
The design polynomials are determined to satisfy the desired
considered N p homogeneous TCP Vegas flows sharing a
specifications in terms of the input tracking, disturbance
bottleneck in their path as in Fig. 1. In order for deriving the
rejection and the noise attenuation, respectively. Usually A(s)
simulation setup parameters, we have examined our campus
contains zeros of the plant (those that are in the left-hand side of
router traffic properties form 9 to 10 AM of Monday, November
the s-plane) and poles of the input disturbance. Table 1 shows
20, 2006. The data have been gathered with Wireshark and at
degree requirements of polynomials A(s) and B(s) for different
the boundary of the campus network resulting in 1 GB of packet
input disturbances [22] in Manabe’s standard form of CDM
headers crossing the router. We have analysed the trace files
design. F(s) is chosen to satisfy the following relation:
to derive three types of data: (1) the average packet length to
derive the bandwidth capacity in terms of packets/s; (2) the

P(s) 
F(s) = . (33) experienced RTT to derive a reasonable value for the round
N (s) s=0
From the above relations, a set of algebraic equations can
be established between the desired characteristic polynomial
coefficients, plant parameters and the CDM controller
parameters. Based on the above discussions, the procedure of
designing of CDM is summarized as [23]
• Choose suitable controller polynomial structures A(s), B(s)
and F(s) based on the disturbance rejection properties,
controller bandwidth and system specifications.
• Choose a proper equivalent time constant τ . The closed loop
settling time is about (2.5–3)τ .
• Choose proper set of the stability indices, γi and the stability
limits, γi∗ .
• Draw the coefficient diagram of the closed loop system
and make some adjustment to satisfy the performance
specification if necessary.
4.2. AQM controller design using CDM
To design AQM controller using CDM, consider the transfer
function of (23). Regarding this transfer function and following
trip propagation delay and round trip time values; (3) the traffic
mixture and flow size distribution to derive an approximate
number of long-lived flows and to validate our assumption
about long-lived flows.
The campus out-link is a 6 Mbps link which is shared
between five local routers allow the campus networks.
Therefore the nominal bandwidth of each local router is
1.2 Mbps. To derive the capacity in terms of packets/s, consider
the observed packet length and packet length distribution shown
in Fig. 8. In order to plot these figures, the pure acknowledge
packets have been omitted. It has been done due to this fact
that there are large numbers of such packets that have length
of less than 80 B. Considering them, however, underestimate
the average packet length and consequently overestimate the
available bandwidth. From this figure, both the mean and the
median of the packet length distribution are 1500 B. Therefore,
the link capacity is calculated as C p0 = 100 packets/s.
The other factor is the experienced RTT. The observed RTT
and its empirical cumulative distribution function have been
shown in Fig. 9. From this figure, the mean RTT value is equal
to 249 ms. therefore, we take d p0 = 0.25 s.
 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155 151

Fig. 8. (a) The empirical observed packet length; and (b) the packet length
distribution. Fig. 10. The composition of IP traffic for different protocols in terms of (a)
number of packets and (b) volume of bytes.

Fig. 11. Experimental flow size in terms of volume of byte versus number of
packets.

Fig. 9. (a) The observed RTT; and (b) its empirical cumulative distribution
function.

Finally, the traffic mixture and the flow distribution versus

number of flows are shown in Figs. 10–12. The composition
of ICMP traffic and IP traffic for TCP, UDP protocols in
terms of packets and bytes are shown in Fig. 10(a) and (b),
respectively. From theses figures, more than 90% of the traffic
comes from TCP protocol. Therefore, in order to investigate the
flow properties, only the TCP flow size distribution has been
considered.
Fig. 11 shows the experimentally observed TCP flow size in
Fig. 12. Histogram of flow size distribution versus (a) volume of bytes,
bytes versus TCP flow size in packets. As it is obvious from this
(b) number of packets.
figure, there are lots of small flows consisting of low number
of packets. Such flows are usually due to Syn attacks, worms 5.2. The simulation results
or spasms and do not carry important data. The flows carrying
data usually have larger number of packets. The approximate To design the controller, we assume the nominal values
distinction point of such flows is shown in Fig. 11 at the mid known to the controller are: N p0 = 55, d p0 = 0.25 s,
point of the graph. This point represents flows with more than C p0 = 100 packets/s, α p = 5 and B p = 50. Based on
100 packets that carry more than 120 KB of data in our network. discussions in Section 3.3, we choose ρ p = 1/4T p∗ = 0.33,
Corresponding points in the histogram of number of flows µ p = 3 and η = 50. Then the network operating point
versus flow length in bytes and packets are shown in Fig. 12. becomes: (q ∗p , x ∗p , pd∗ ) = (50, 1.818, 0.0616) and the small
From these plots, the average number of active flows carrying signal transfer function of system is:
data is about 55. Based on this discussion we take the nominal
6173
number of active flows N p0 = 55. G n (s) = . (35)
0.75s 3 + 42.69s 2 + 14.15s + 20.58
152 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155
Fig. 13. CDM controller performance for in comparison with PI and RED for
nominal parameters: CDM (solid), PI (dash), and RED (dot).
To design the CDM controller, we should choose proper
values for τ and γi as stated in Section 4.2. Selecting τ =
4T p∗ = 3 s and γi = [∞, 7.5, 6, 6, 6, 6, ∞], the closed loop
characteristic polynomial becomes P(s) = 3.134e−6s 6 +
0.0102s 5 + 5.4874s 4 + 493.86s 3 + 7.407e3s 2 + 1.9e4s +
6.173e3. Therefore, the controller polynomials are:
A(s) = 4.18e−6s 3 + 0.0133s 2 + 6.56s
(36)
B(s) = 0.0346s 3 + 1.1849s 2 + 2.978s + 1.
For comparison, two well-known AQM controllers i.e.
RED [6,10] and proportional plus integral (PI) controllers
are also simulated. The RED parameters are set the same
as those of [6]. The parameters maximum drop probability,
minimum queue threshold and maximum queue threshold
(i.e. pmax , minth and maxth ) of RED are set to 0.1, 10, and
150, respectively. The PI controller has been designed based on
Ziegler–Nichols method [28]. The resulting parameters become
k p = 0.0572 and ki = 0.0475 as the proportional and integral
gains, respectively. The responses of the designed controllers
for the nominal network parameters are shown in Fig. 13.
From the graphs in this figure, it is observed that both queue
and rate regulation has been best done by the CDM method.
PI performance is also good in this case, but, the transient
behaviour of CDM is better than PI. Damping oscillations in
response and a large settling time as well as a small steady-state
error of the queue length are the properties of RED performance
with respect to CDM and PI AQM controllers.
In continue we investigate the performance of the
designed controllers under different unknown and time-varying
circumstances.
Fig. 14. CDM controller performance for in comparison with PI and RED for
unknown parameters: CDM (solid), PI (dash), and RED (dot).
5.2.1. Exp. 1: Unknown parameters
The performance of the designed CDM controller along
with those of RED and PI controllers are investigated for the
case that the network parameters are assumed unknown. It is
assumed that the real values of the plant parameters are: N p =
65, d p = 0.35 s, and C p = 80 packets/s and all other network
and controller parameters remain the same as considered for
the nominal network. The simulation result for regulating the
queue length around B p = 50 has been shown in Fig. 14. It
is seen that the queue regulation objective has been obtained
for CDM, while for PI some small oscillations are observed
in this case. For RED, the equilibrium queue length has been
doubled, which is due to the RED algorithm which does not
explicitly consider the desired queue length for calculating
packet marking/dropping probability [28]. Although the packet
marking probability of CDM has been increased due to the
difference in the operating point in effect of the network
parameter mismatches, it is still less than that PI controller.
However, RED controller has the smallest packet marking
probability among all.
Seeking for the limits of applicability, the behaviour of the
designed controllers for various load/network parameters has
been investigated in the above described network. For this
purpose, ±50% variations have been applied to the parameters
C p , N p and d p ; and the controllers’ performance have been
examined. The underlying results have been summarized in
Table 2. From this table, the simulation results show that
the CDM controller keeps its performance for all over the
variation ranges. The steady state queue length does never
show deviation more than 12% from its desired value and the
response settles down in a reasonable time with respect to the
nominal response. Moreover, such performance is observed
for larger ranges of parameter variations as well (up to 100%
 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155 153

Table 2
The performance summary of the designed controllers for ±50% variations in load/network parameters

Parameter Stabilty region (%) Steady state queue vaiations in stable region Setting time variations
CDM PI RED CDM PI RED CDM PI RED
Np ±50% [−5%, 40%] [−20%, 50%] ±6% ±24% [−20%, 70%] ±15% [−100%, 200%] ±40%
Asc. Desc. Asc. Desc. Desc. Desc.
Cp ±50% [−50%, 5%] [−50%, 30%] ±5% ±25% [−20%, 10%] −25%, 10% [+15%, 40%] ±50%
Asc. Asc. Asc. Desc. Asc. Desc.
dp ±50% [−10%, 50%] [−30%, 50%] [−15%, +4%] ±50% ±25%] [−6%, 14%] [−20%, 8%] ±35%
Asc. Desc. Asc. Asc. Desc. Asc.
Here ‘Asc.’ abbreviates ‘Ascending’ and ‘Desc.’ abbreviates ‘Descending’.

Fig. 15. CDM controller performance for in comparison with PI and RED for time-varying parameters: (a) CDM (left column graphs), (b) PI (middle graphs), and
(c) RED (right column graphs).

increase). On the other hand, PI and RED become unstable for
parameter variations outside the specified ranges in Table 2.
Besides, the ability of queue length regulation and the transient
response of PI and RED degrade dramatically in comparison
with CDM. For the network with PI as AQM controller,
deviation from nominal parameters leads to small oscillations
around the queue length equilibrium point. The amplitude
of such oscillations increases as the deviation increases until
it reaches instability; while the margins of stability are not
symmetrical with respect to the nominal parameters. Queue
length equilibrium point in a network employing RED, changes
in a linear ascending/descending manner as mentioned in
Table 2. Besides, for RED and PI, the settling time varies
significantly in effect of parameter variations which is not
desirable.
5.2.2. Exp. 2: Time-varying parameters
In Fig. 15, the performance of the CDM controller is
compared with the PI and RED AQM controllers in time-
varying situation. For this purpose, step changes in the
parameters C p , N p and d p have been applied to the network
according to the last row of Fig. 15. As seen, while both PI and
154 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155
Fig. 16. CDM robustness index ratios γi /γi∗ (left column graphs) and CDM instability indices γi γi+1 (right column graphs) for various network parameter ratios;
(a) and (b) for N p /N p0 , (c) and (d) for C p /C p0 , (e) and (f) for d p /d p0 .
RED show oscillating behaviours in response to such variations,
by employing CDM controller both data sending rate and queue
length have been stabilized around their desired values with a
relatively low marking/dropping probability.
5.2.3. Exp. 3: Robustness analysis
As discussed in Section 4, robustness with respect to
network parameter variations is one of the major goals in
AQM design. For CDM as an AQM method, the stability
and instability indices have been introduced in (30a) and
(30b). Besides, robustness condition has been given in (30c).
Regarding this relationships, the robustness and stability
properties can be investigated through investigating the index
ratios of γi /γi∗ for i = 2, 3, 4 and instability indices of γi γi+1
for i = 2, 3, 4. The graphs in Fig. 16 show the variations of
these index ratios and instability indices for different values of
network parameters i.e. N p , C p , and d p . From these figures,
increase in N p and d p increases the robustness and stability
of the system while increase in C p alleviates both of them.
However, the index ratios and the instability indices remain
more than 5 and 10 for all the investigated range for C p which is
representative of very good margins of stability and robustness.
These values are pretty more for N p and d p as observed in the
Fig. 16, which guarantee good performance of the controller in
a dynamically varying network. Note however that these results
are in agreement with the simulation results in Exp. 1.
6. Conclusion
In this paper, a mathematical model and control theoretical
framework for designing AQM controllers in networks
supporting TCP Vegas is introduced. We have emphasized on a
modified TCP Vegas algorithm that can respond to congestion
signals through ECN. By this property, Vegas achieves ability
of cooperating with AQM algorithms employed in routers to
support the QoS requirements. In this paper, at first the overall
nonlinear delayed differential equations of the dynamics model
of closed loop system have been derived based on TCP Vegas
model. The model is then linearized to derive a transfer function
representation between the packet marking probability and the
bottleneck router queue length as the input and output of the
modified TCP Vegas/AQM system. For the case of a network
with homogeneous data sources sharing a single bottleneck,
an AQM controller has been designed based on Coefficient
Diagram Method (CDM). The controller has been previously
employed by the authors as an AQM method for networks
supporting TCP Reno and its out-performance with respect to
the other controllers such as RED and PI has been shown. This
good background motivated the authors to design the CDM
as AQM controller for networks supporting TCP Vegas. In
this paper, the performance of the TCP Vegas/CDM networks
has been evaluated through different simulation scenarios
and its out-performance with respect to other AQM methods
i.e. RED and PI (that are simulated for comparison) has been
shown. It should be noted that designing any model-based
AQM controller, and especially CDM, is strongly correlated
with TCP variant and its model. For example, the degree
of CDM polynomials and the selected CDM stability indices
for a network with the same load, bandwidth, and delay
parameters are not the same for networks supporting TCP
Vegas and networks supporting TCP Reno. However, the good
performance of CDM with both types of TCPs is representative
of the good potential of CDM to be employed as an AQM
controller.
In order to synthesize our simulation scenario, the properties
of the traffic passing our campus router have been measured
through some experiments. Packet length and round trip time
(RTT), traffic composition, and flow size and number of active
 N. Bigdeli, M. Haeri / ISA Transactions 47 (2008) 143–155
flows are the considered features of the measured data. We have
analysed the trace files to derive three types of data: (1) the
average packet length to determine the bandwidth capacity in
terms of packets/s; (2) the experienced RTT to determine a
reasonable value for the round trip propagation delay and round
trip time values; (3) the traffic mixture and flow size distribution
to determine an approximate number of active long-lived flows
and to validate our assumption about long-lived flows. Our
analysis results show that more than 90% of passing IP traffic
comes from TCP sources while a low percent of TCP flows
really carry data. However, these flows called long-live flows
are dominant in flow size both in number of packets and volume
of bytes.
Simulation results indicate that the CDM-AQM controller
performs very well for network supporting TCP Vegas even in
the case of unknown network parameters or parameter varia-
tions. Simulation results illustrate capability of the controller to
cope with the network variations. As another point, the stability
and robustness properties of the designed controller have been
discussed in this paper that can help the designer to have a judg-
ment about the performance margins of the designed controller.
Acknowledgements
Authors gratefully acknowledge the financial support of
the Iranian Telecommunication Research Centre (ITRC) under
Grant T500-8210. The authors also deeply appreciate the
Information Technology Services centre of Sharif University
of Technology for supporting them with the measurement
facilities.
References
[1] Moronski J. Freeze TCP: A true end-to-end TCP enhancement mechanism
for mobile environment. In: Proc. of IEEE INFOCOM 2000. p. 1537–45.
[2] Brakmo LS, Peterson LL. TCP Vegas: end to end congestion avoidance
on a global internet. IEEE Journal on Selected Areas on Communications
1995;13(8):1465–80.
[3] Bonald T. Comparison of TCP Reno and TCP Vegas via fluid
approximation. Tech. rep. RR 3563. 1998.
[4] Mo J, La RJ, Anantharam RJ, Walrand JC. Analysis and comparison of
TCP Reno and Vegas. In: Proc. of IEEE INFOCOM 1999. p. 1556–63.
[5] Low SH, Peterson LL, Wang L. Understanding TCP Vegas: A duality
model. Journal of ACM 2002;49(2):207–35.
[6] Kim K, Choi CH. Queue delay estimation and its application to TCP
Vegas. Computer Networks 2003;43(5):619–31.
[7] Haeri M, Mohsenian Rad AH. Adaptive model predictive TCP delay-
155
based congestion control. Computer Communications 2006;29(11):
1963–78.
[8] Srijith KN, Jacob L, Ananda AL. TCP Vegas-A: Improving the
performance of TCP Vegas. Computer Communications 2005;28(4):
429–40.
[9] Athuraliya S, Li VH, Low SH, Yin Q. REM: Active queue management.
IEEE Network 2001;15(3):48–53.
[10] Floyd S, Jacobson V. Random early detection gateways for congestion
avoidance. IEEE/ACM Transactions on Networking 1993;1:397–413.
[11] Misra V, Gong WB, Towsley D. Fluid-based analysis of a network of
AQM routers supporting TCP flows with an application to RED. In: Proc.
ACM/SIGCOMM 2000. p. 151–61.
[12] Hollot CV, Misra V, Towsley D, Gong WB. Analysis and design of
controllers for AQM routers supporting TCP flows. IEEE Transactions
on Automatic Control 2002;47:945–59.
[13] Chang X, Muppala JK. A stable queue-based adaptive controller
for improving AQM performance. Computer Networks 2006;50(13):
2204–24.
[14] Bigdeli N, Haeri M. Design of a robust AQM strategy for dynamic
TCP/AQM networks based on CDM. In: Proc. of CCA 2005. p. 716–21.
[15] Bigdeli N, Haeri M. Adaptive-CDM: A new AQM controller to cope with
TCP/AQM networks dynamics. In: Proc. of ICICS 2005. p. 1520–4.
[16] Bigdeli N, Haeri M. Performance analysis of CDM as an AQM congestion
controller in dynamic networks. In: Proc. of EuroCon 2005. p. 680–83.
[17] Ramakrishnan K, Floyd S. A proposal to add explicit congestion
notification. Tech. rep. RFC, 2481. 1998.
[18] Floyd S, Kohler E. Internet research needs better models. In: Proc. of first
workshop on hot topics in networks (HotNets-I) 2002.
[19] Choe H, Low SH. Stabilized vegas. In: Proc. of IEEE INFOCOM 2003.
p. 2290–300.
[20] Manabe S. The Coefficient diagram method. In: Proc. of the 14th IFAC
symposium on automatic control in aerospace 1998. p. 199–210.
[21] Manabe S. Coefficient diagram method as applied to attitude control
of controlled-bias-satellite. In: Proc. of the 13th IFAC symposium on
automatic control in aerospace 1994. p. 322–7.
[22] Manabe S. Brief tutorial and survey of coefficient diagram method. In:
Proc. of the 4th Asian control conference proceedings 2002. p. 1161–6.
[23] Hamamci SE. Simple polynomial controller design by the coefficient
diagram method. WSEAS Transactions on Circuits and Systems 2004;
3(4):951–6.
[24] Manabe S. Controller design of two-mass resonant system by coefficient
diagram method. The Transactions of The Institute of Electrical Engineers
of Japan 1998;118-D: 1(1):58–66.
[25] Lipatov A, Sokhlov N. Some sufficient conditions for stability and
instability of continuous linear stationary systems. Automatic Remote
Control 1978;39:1285–91.
[26] Hamamci SE, Ucar A. A robust model-based control for uncertain
systems. Transaction Institute of Measurement and Control 2002;24(4):
331–45.
[27] Hamamci SE. A robust polynomial-based control for stable processes with
time delay. Electrical Engineering 2005;87(3):163–72.
[28] Goodwin CG, Graebe SF, Salgado ME. Control System Design. New
Jersey (USA): Prentice-Hall, Inc.; 2001.