University of Pennsylvania

ScholarlyCommons
Departmental Papers (ESE) Department of Electrical & Systems Engineering

April 2004

An Adaptive Virtual Queue (AVQ) Algorithm for

Active Queue Management
Srisankar S. Kunniyur
University of Pennsylvania, kunniyur@[Link]

R. Srikant
University of Illinois

Follow this and additional works at: [Link]

Recommended Citation
Srisankar S. Kunniyur and R. Srikant, "An Adaptive Virtual Queue (AVQ) Algorithm for Active Queue Management", . April 2004.

Copyright 2004 IEEE. Reprinted from IEEE/ACM Transactions on Networking, Volume 12, Issue 2, April 2004, pages 286-299.
Publisher URL: [Link]

This material is posted here with permission of the IEEE. Such permission of the IEEE does not in any way imply IEEE endorsement of any of the
University of Pennsylvania's products or services. Internal or personal use of this material is permitted. However, permission to reprint/republish this
material for advertising or promotional purposes or for creating new collective works for resale or redistribution must be obtained from the IEEE by
writing to pubs-permissions@[Link]. By choosing to view this document, you agree to all provisions of the copyright laws protecting it.

This paper is posted at ScholarlyCommons. [Link]

For more information, please contact repository@[Link].
An Adaptive Virtual Queue (AVQ) Algorithm for Active Queue
Management
Abstract
Virtual queue-based marking schemes have been recently proposed for Active Queue Management (AQM) in
Internet routers. We consider a particular scheme, which we call the Adaptive Virtual Queue (AVQ), and
study its following properties: its stability in the presence of feedback delays, its ability to maintain small
queue lengths, and its robustness in the presence of extremely short flows (the so-called web mice). Using a
linearized model of the system dynamics, we present a simple rule to design the parameters of the AVQ
algorithm. We then compare its performance through simulation with several well-known AQM schemes such
as RED, REM, Proportional Integral (PI) controller, and a nonadaptive virtual queue algorithm. With a view
toward implementation, we show that AVQ can be implemented as a simple token bucket using only a few
lines of code.

Keywords
Active queue management (AQM), ECN marking, Internet congestion control

Comments
Copyright 2004 IEEE. Reprinted from IEEE/ACM Transactions on Networking, Volume 12, Issue 2, April
2004, pages 286-299.
Publisher URL: [Link]

This material is posted here with permission of the IEEE. Such permission of the IEEE does not in any way
imply IEEE endorsement of any of the University of Pennsylvania's products or services. Internal or personal
use of this material is permitted. However, permission to reprint/republish this material for advertising or
promotional purposes or for creating new collective works for resale or redistribution must be obtained from
the IEEE by writing to pubs-permissions@[Link]. By choosing to view this document, you agree to all
provisions of the copyright laws protecting it.

This journal article is available at ScholarlyCommons: [Link]

286 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 12, NO. 2, APRIL 2004

An Adaptive Virtual Queue (AVQ) Algorithm for

Active Queue Management
Srisankar S. Kunniyur, Member, IEEE, and R. Srikant, Senior Member, IEEE

Abstract—Virtual queue-based marking schemes have been re- packets to be marked in a manner that conveys information
cently proposed for Active Queue Management (AQM) in Internet about the current state of the network to the users. Algorithms
routers. We consider a particular scheme, which we call the Adap- that the routers employ to convey such information are called
tive Virtual Queue (AVQ), and study its following properties: its
stability in the presence of feedback delays, its ability to maintain Active Queue Management (AQM) schemes. An AQM scheme
small queue lengths, and its robustness in the presence of extremely might mark or drop packets depending on the policy at the
short flows (the so-called web mice). Using a linearized model of router. In this paper, we use the term “marking” more generally
the system dynamics, we present a simple rule to design the pa- to refer to any action taken by the router to notify the user of
rameters of the AVQ algorithm. We then compare its performance incipient congestion. The action can, in reality, be ECN-type
through simulation with several well-known AQM schemes such as
RED, REM, Proportional Integral (PI) controller, and a nonadap- marking or dropping (as in RED) depending upon the policy
tive virtual queue algorithm. With a view toward implementation, set for the router. As in earlier work on studying AQM schemes
we show that AVQ can be implemented as a simple token bucket [6], [7], [15], this distinction is blurred in the mathematical
using only a few lines of code. analysis to allow for the development of simple design rules
Index Terms—Active queue management (AQM), ECN for the choice of AQM parameters. However, our simulations
marking, Internet congestion control. consider marking and dropping schemes separately.
Designing robust AQM schemes has been a very active
research area in the Internet community. Some AQM schemes
I. INTRODUCTION
that have been proposed include RED [4], a virtual queue-based

I N the modern day Internet, there has been a strong demand

for quality of service (QoS) and fairness among flows. As
a result, in addition to the sources, the links are also forced to
scheme where the virtual capacity is adapted [10], [11], [13],
SRED [16], Blue [2], Proportional Integral (PI) controller [7],
REM [1], a virtual queue based AQM scheme [5] (which we
play an active role in congestion control and avoidance. Random refer to as the Gibbens–Kelly Virtual Queue, or the GKVQ
Early Discard (RED) [4] was originally proposed to achieve scheme) among others. While most of the proposed AQM
fairness among sources with different burstiness and to control schemes detect congestion based on the queue lengths at
queue lengths. RED allows for dropping packets before buffer the link (e.g., RED), some AQM schemes detect congestion
overflow. Another form of congestion notification that has been based on the arrival rate of the packets at the link (e.g., virtual
discussed since the advent of RED is Explicit Congestion Noti- queue-based schemes) and some use a combination of both
fication (ECN) [3]. ECN has been proposed to allow each link (e.g., PI). Also, most of the AQM schemes involve adapting
to participate in congestion control by notifying users when it the marking probability (as noted before, we use the term
detects an onset of congestion. Upon detecting incipient conges- marking to refer to both marking and dropping) in some way
tion, a bit in the packet header is set to one for the purpose of or another. An important question is how fast should one adapt
notifying the user that a link on its route is experiencing conges- while maintaining the stability of the system? Here the system
tion. The user then reacts to the mark as if a packet has been lost. refers jointly to the TCP congestion controllers operating at
Thus, the link avoids dropping the packet (thereby enhancing the edges of the network and the AQM schemes operating in
goodput) and still manages to convey congestion information to the interior of the network. Adapting too fast might make the
the user. system respond quickly to changing network conditions, but
To provide ECN marks or drop packets in order to control it might lead to large oscillatory behavior or in the worst-case
queue lengths or provide fairness, the routers have to select even instability. Adapting too slowly might lead to sluggish
behavior and more losses or marks than desired, which might
lead to a lower throughput.
Manuscript received August 3, 2001; revised May 6, 2002; approved by In this paper, we start by presenting an implementation of a
IEEE/ACM TRANSACTIONS ON NETWORKING Editor J.-Y. LeBoudec. This
work was supported by the Defense Advanced Research Projects Agency virtual-queue based AQM scheme, namely the Adaptive Virtual
(DARPA) under Grant F30602-00-2-0542 and by the National Science Queue (AVQ). The motivation behind the AVQ algorithm is to
Foundation under Grants ANI-9813710 and NCR-9701525. design an AQM scheme that results in a low-loss, low-delay and
S. S. Kunniyur is with the Department of Electrical and Systems Engi-
neering, University of Pennsylvania, Philadelphia, PA 19104 USA (e-mail: high utilization operation at the link. We then discuss a method-
kunniyur@[Link]). ology for finding the fastest rate at which the marking proba-
R. Srikant is with the Department of Electrical and Computer Engineering and bility adaptation can take place, given certain system parame-
Coordinated Science Laboratory, University of Illinois at Urbana-Champaign,
Urbana, IL 61801 USA (e-mail: rsrikant@[Link]). ters like the maximum delay and the number of users, so that
Digital Object Identifier 10.1109/TNET.2004.826291 the system remains stable. We note that the marking probability
1063-6692/04$20.00 © 2004 IEEE
KUNNIYUR AND SRIKANT: AVQ ALGORITHM FOR ACTIVE QUEUE MANAGEMENT 287

in AVQ is implicit, no marking probability is explicitly calcu-

lated and thus, no random number generation is required. On the
other hand, we replace the marking probability calculation with
the computation of the capacity of a virtual queue. Motivated by
the success of the analysis and design of other AQM schemes in
[6], [7], and [15], we consider a single router accessed by many
TCP sources with the same round-trip time (RTT) and use a con-
trol-theoretic analysis to study the stability of this system.
The AVQ algorithm maintains a virtual queue whose capacity
(called virtual capacity) is less than the actual capacity of the
Fig. 1. AVQ: when virtual queue is not full.
link. When a packet arrives in the real queue, the virtual queue
is also updated to reflect the new arrival. Packets in the real
queue are marked/dropped when the virtual buffer overflows.
The virtual capacity at each link is then adapted to ensure that
the total flow entering each link achieves a desired utilization
of the link. This was originally proposed in [10] and [11] as
a rate-based marking scheme. In the absence of feedback de-
lays, it was shown in [13] that a fluid-model representation of
the above scheme, along with the congestion-controllers at the
end-hosts, was semi-globally asymptotically stable when the
link adaptation is sufficiently slow. An appealing feature of the
AVQ scheme is that, in the absence of feedback delays, the
Fig. 2. AVQ: when the incoming packet is dropped from the virtual queue.
system is fair in the sense that it maximizes the sum of utili-
ties of all the users in the network [13]. Combining this with
a result in [10] and [11] which shows that a TCP user with an stability, the scheme exhibits sluggishness when short flows
RTT of can be approximated by a user with a utility function are introduced. This is the subject of simulations in Section III.
, where is the rate of the TCP user, shows that the Since the analytical model does not capture the discrete packet
network as a whole converges to an operating point that min- behavior of the network in addition to the slow-start and the
imizes . This utility function called the poten- timeout characteristics of TCP, the simulations in Section III are
tial delay was introduced as a possible fairness criterion in [14]. intended to demonstrate the performance of the AVQ algorithm
The throughput under this utility function is given by , under these realistic conditions. In Section IV, we provide
where is the loss probability seen by User which is consis- theoretical justification for the design rules in Section II. The
tent with the models in [6], [7], and [15]. While we use this sim- design rules in Section IV are less restrictive than the design
plified model for analysis in the paper, our simulations in ns-2 rules in [12]. Conclusions are provided in Section V.
use TCP-Reno, including slow-start, timeout, fast retransmit,
etc. A slightly more refined utility function is used in [8] and II. AVQ ALGORITHM
the results in this paper can be easily modified to incorporate
that utility function. Let be the capacity of a link and be the desired utilization
We start with a fluid-model representation of the TCP at the link. The AVQ scheme, as presented in [10], [11], and
flow-control, along with the AVQ scheme that was proposed [13], at a router works as follows.
in [10] and [11]. However, here we explicitly consider the • The router maintains a virtual queue whose capacity
feedback delay due to the RTT of each user and thus, we and whose buffer size is equal to the buffer size of the
obtain a set of delay-differential equations. We linearize this real queue. Upon each packet arrival, a fictitious packet is
system and obtain conditions for local stability in terms of the enqueued in the virtual queue if there is sufficient space in
round-trip delay, the number of users, the utilization of the the buffer (see Fig. 1). If the new packet overflows the vir-
link and a smoothing parameter in the update equation of the tual buffer, then the packet is discarded in the virtual buffer
AVQ scheme. The rest of the paper is organized as follows: in and the real packet is marked by setting its ECN bit or the
Section II, we present an implementation of the AVQ algorithm real packet is dropped, depending upon the congestion no-
and provide design rules for the stability of the AVQ and TCP tification mechanism used by the router (see Fig. 2).
together. In Section III, we provide detailed ns-2 simulations • At each packet arrival epoch, the virtual queue capacity is
to validate our design rules and also compare the AVQ algo- updated according to the following differential equation:
rithm with RED, REM, GKVQ and the PI controllers. The
PI controller is somewhat similar to AVQ in that it adapts the (1)
marking probability in a manner similar to the virtual capacity
adaptation in the AVQ scheme, but it depends on the queue size where is the arrival rate at the link and is
at the link. As a result, for small buffers the system tends to the smoothing parameter. The rationale behind this is
perform poorly. Also, since the marking probability is directly that marking has to be more aggressive when the link
modified and this update has to be slow enough for system utilization exceeds the desired utilization and should be
288 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 12, NO. 2, APRIL 2004

3) Instead of attempting to regulate queue length as in RED,

PI controller, or recent versions of REM, we regulate uti-
lization. As we will see in simulations, this is more robust
to the presence of extremely short flows or variability in
the number of long flows in the network. The reason is
that, when utilization is equal to one, variance introduced
by the short flows seems to lead to an undesirable tran-
sient behavior where excessively large queue lengths per-
sist over long periods of time.
4) Unlike the GKVQ algorithm [5], we adapt the capacity of
the virtual queue. A fixed value of leads to a utilization
Fig. 3. AVQ: token bucket implementation of the virtual capacity adaptation.
that is always smaller than and it could be much
smaller than this depending on the number of users in
less aggressive when the link utilization is below the the system. Our marking mechanism is also different in
desired utilization. that we do not mark until the end of a busy period after a
We now make the following observations. No actual enqueueing congestion episode.
or dequeueing of packets is necessary in the virtual queue, we 5) There are two parameters that have to be chosen to im-
just have to keep track of the virtual queue length. Equation (1) plement AVQ: the desired utilization and the damping
can be thought of as a token bucket where tokens are generated factor . The desired utilization determines the robust-
at rate up to a maximum of and depleted by each arrival ness to the presence of uncontrollable short flows. It al-
by an amount equal to times the size of the packet (see Fig. 3). lows an ISP to trade-off between high levels of utilization
Define: and small queue lengths. Both the parameters and de-
buffer size; termine the stability of the AVQ algorithm and we provide
arrival time of previous packet; a simple design rule to choose these parameters.
current time; The starting point for the analysis of such a scheme is the fluid
number of bytes in current packet; model of the TCP congestion-avoidance algorithm as proposed
number of bytes currently in the virtual queue. in [10] and [11]. A theoretical justification of how a stochastic
Then, the following pseudo-code describes an implementa- discrete-time equation can be approximated by a fluid-model is
tion of AVQ scheme. shown in [18] and [19]. We then incorporate the virtual capacity
update equation with this model and study the stability of the
entire system under linearization.
The AVQ Algorithm
Consider a single link of capacity with TCP users
At each packet arrival epoch do
V Q max(V Q 0 C~ (t 0 s); 0) /3Update Virtual traversing it. Let the desired utilization of the link be ,
Queue Size 3/
and let be the round-trip propagation delay of each user (we
assume that all users have the same round-trip propagation
If V Q + b > B
delay). Let be the flow rate of user at time . We will
Mark or drop packet in the real queue
else
model the TCP users using the utility function as
V Q V Q + b /3Update Virtual Queue proposed in [10] and [11]. For the sake of simplicity and
Size 3/
tractability, we will neglect slow-start and timeouts when mod-
eling the TCP users. We will later show through simulations
endif
C~ = max(min(C~ +  3 3 C 3 (t 0 s); C ) 0  3 b; 0) /3 that even with slow-start and timeouts, stability is maintained.
Update Virtual Capacity 3/
Let be the fraction of packets marked at the link. The
s t /3Update last packet arrival time fraction of packets marked (i.e., ) is a function of the total
3/ arrival rate at the link as well as the virtual capacity of the link.
The congestion-avoidance algorithm of TCP user can now be
represented by the following delay differential equation:
We note the following features of the AVQ scheme.
1) The implementation complexity of the AVQ scheme is
comparable to RED. RED performs averaging of the (2)
queue length, dropping probability computation and
random number generation to make drop decisions. We
replace these with the virtual capacity calculation in where and is the virtual-capacity of the link. A
AVQ. value of 2/3 would give us the steady-state throughput of TCP as
2) AVQ is a primarily a rate-based marking, as opposed , where is the steady-state marking probability
to queue length or average queue length-based marking. which is consistent with the results in [17]. Hence, we will use
This provides early feedback, the advantages of which , in all our calculations. Also, note that on substituting
have been explored by Hollot et al. [6], [7], which was , where is the window-size of user , we recover
also mentioned in Kelly et al. [9]. the TCP window control algorithm [10], [11], [15].
KUNNIYUR AND SRIKANT: AVQ ALGORITHM FOR ACTIVE QUEUE MANAGEMENT 289

The update equation at each link can now be written as III. SIMULATIONS
The above theorem shows that by choosing according
(3)
to (7), one can guarantee the local stability of the TCP/AQM
where is the total flow into the link and scheme. However, the fluid model does not take into account
is the smoothing parameter. Note that determines how fast the discrete packet behavior of the network as well as the
one adapts the marking probability at the link to the changing inherent nonlinearities in the TCP algorithm. As a result,
network conditions. We will present a design rule that specifies it becomes important to verify the analytical results using
how to choose for a given feedback delay , utilization , and simulations in which the nonlinearities of the system are taken
a lower bound on the number of users . In fact, as we will into account. In this section, we conduct experiments that
show in Section IV, one can derive bounds on any of the four simulate various scenarios in the network and show that the
parameters , , , or , given the other three using the same AVQ algorithm performs as predicted by the analytical model
design rule. However, in practice, it would seem most natural to in all the experiments. Even though each experiment shows that
choose given the other three parameters. AVQ results in small queues, low loss and high utilization at
Let , , , and denote the equilibrium values of the link, it is important to note that each experiment simulates
, , and . The equilibrium point of the nonlinear a different scenario and the performance of AVQ is tested in
TCP/AQM model is given by this scenario.
In this section, we use the packet simulator ns-2 1 to simu-
late the AVQ scheme. We show that the simulation results agree
with the convergence results shown in the previous section. In
particular, we select an , using Theorem 2.1 that will ensure
stability for a given round-trip delay and a lower bound on the
number of users, . We then present a single set out of many
experiments that we did to show that indeed stabilizes the
system even in the presence of arrivals and departure of short
Let us assume that connections. We then compare this scheme with many other
AQM schemes.

A. Simulation Setup
The linearized version of the nonlinear TCP/AQM model can Throughout this section, we consider a single link of capacity
now be written as 10 Mb/s that marks or drops packets according to some AQM
scheme. For AVQ, unless otherwise stated, we let , the desired
(4) utilization, be 0.98. We use TCP-Reno as the default transport
(5) protocol with the TCP data packet size set to 1000 bytes. Each
TCP connection is placed in one of five classes which differ only
where in their round-trip propagation delays. Class 1 has a round-trip
delay of 40 ms, Class 2 has a round-trip delay of 60 ms, Class 3
has a round-trip delay of 80 ms, Class 4 has a round-trip delay
of 100 ms, and Class 5 has a round-trip delay of 130 ms. The
and buffer size at the link is assumed to be 100 packets.
In the first five experiments, we assume that the link marks
packets and thus, any packet loss is due to buffer overflow.
In these experiments, we demonstrate that the AVQ scheme
achieves high utilization and low packet loss. Further, the al-
For analytical tractability, we assume that gorithm responds quickly to changing network conditions such
as varying number of TCP flows. In the first experiment, we
study the convergence properties of the AVQ algorithm both in
(6) the absence and in the presence of short flows. In the remaining
experiments, we compare the AVQ algorithm with other AQM
We will now state the main result of this paper which serves schemes like RED, REM, PI, and GKVQ. In the second experi-
as the design for the AVQ algorithm. The proof of this result is ment, we compare the performance of various AQM schemes
given in Section IV. in the presence of long-lived flows. In the third experiment,
Theorem 2.1: Suppose that the feedback delay , number of we compare the transient behavior of the AQM schemes when
users , and the utilization , are given. Let be given by long-lived flows are dropped and added to the network. In the
fourth experiment, we compare the AQM schemes in the pres-
(7) ence of short flows in the network. In the fifth experiment, we
study the sensitivity of the AVQ algorithm to the smoothing
Then, for all , the system is locally stable. 1[Online.] Available: [Link]
290 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 12, NO. 2, APRIL 2004

Fig. 4. Experiment 1. Evolution of the virtual capacity with time for the AVQ Fig. 5. Experiment 1. Queue length versus time for the AVQ scheme.
scheme.
TABLE I
EXPERIMENT 1. MEAN AND THE STANDARD DEVIATION OF THE QUEUE SIZE
parameter as well as to the desired utilization parameter . BEFORE AND AFTER THE INTRODUCTION OF SHORT FLOWS
In the last experiment, we compare the AVQ scheme with other
schemes when the link drops packets (as opposed to marking) to
indicate congestion. Again, the AVQ scheme is shown to have
smaller queue lengths compared to other schemes.
We design the AVQ controller for a maximum delay of
ms. Using the design rule in Theorem 2.1, any , will
ensure stability. In the experiments, to account for nonlinearities due to the addition of short flows in the system. Another impor-
in the system, we let be 0.15. In all experiments, we consider tant performance measure is the number of packets dropped due
two types of flows: FTP flows that persist throughout the du- to buffer overflow in the system. Since ECN marking is used,
ration of the simulations and FTP flows of 20 packets each (to we expect the number of packets lost due to buffer overflow
model the short flows). to be small. Indeed, only 10 out of roughly 250 000 packets are
Experiment 1: In this experiment, we study the convergence dropped. These drops are primarily due to the sudden additional
properties and buffer sizes at the queue for the AVQ scheme load brought on by the short flows. Another performance mea-
alone. A total of 180 FTP flows with 36 in each delay class per- sure that is of interest is the utilization of the link. The utilization
sist throughout the duration of the simulations, while short flows was observed to be 0.9827, which is very close to the desired uti-
(of 20 packets each) arrive at the link at the rate of 30 flows lization of 0.98. We note that the apparent discrepancy between
per second. The short flows are uniformly distributed among Fig. 5, where the queue length never reaches the buffer size of
the five delay classes. To simulate a sudden change in network 100 packets, and the fact that there are ten dropped packets is
conditions, we start the experiment with only FTP flows in the due to the fact that the queue length is sampled only once every
system and introduce the short flows after 100 s. The evolution 100 ms to plot the graphs.
of the virtual capacity is given in Fig. 4. After an initial transient, We will now compare the AVQ scheme with other AQM
the virtual capacity settles down and oscillates around a partic- schemes that have been proposed. Since there are many AQM
ular value. Note that the oscillations in the virtual capacity are schemes in the literature, we will compare the AVQ scheme with
due to the packet nature of the network which is not captured a representative few. In particular, we will compare the AVQ
by the analytical model. At 100 s, there is a drop in the vir- scheme with:
tual capacity since the AVQ algorithm adapts to the changing 1) Random Early Discard (RED) proposed in [4]. In our ex-
number of flows. Beyond 100 s, the virtual capacity is lower periments, we use the “gentle” version of RED. Unless
than it was before 100 s since the links marks packets aggres- otherwise stated, the parameters were chosen as recom-
sively due to the increased load. The queue length evolution for mended in [20].
the system every 100 ms is given in Fig. 5. Except during tran- 2) Random Early Marking (REM) proposed in [1]. The
sients introduced by load changes, the queue lengths are small REM scheme tries to regulate the queue length to a
(less than 20 packets). At 100 s, the queue length jumps up due desired value (denoted by ) by adapting the marking
to the short flows. However, the system stabilizes and the queue probability. The REM controller marks each packet with
lengths are small once again. Table I gives the average and the a probability which is updated periodically (say, every
standard deviation of the queue length before and after the in- seconds) as
troduction of short flows. Note that there is a small increase in
the average queue length as well as in the standard deviation
KUNNIYUR AND SRIKANT: AVQ ALGORITHM FOR ACTIVE QUEUE MANAGEMENT 291

where is a arbitrary constant greater than one and

and and are constants and is the queue length

at the sampling instant. Since REM is very sensitive
to , we will use the values as recommended in [1].
3) The PI controller proposed in [6]. The PI controller marks
each packet with a probability which is updated period-
ically (say, every seconds) as

where and are constants chosen according

to the design rules given in [6].
4) The virtual-queue-based AQM scheme (GKVQ) pro-
posed in [5]. In this scheme, the link maintains a virtual
queue with fixed capacity , and buffer size
, where , and is the buffer capacity
of the original queue. Whenever the virtual queue
overflows, all packets in the real queue and all future Fig. 6. Experiment 2. Losses at the link for varying number of FTP
incoming packets are marked until the virtual queue connections for the different AQM schemes.
becomes empty again. Note that this scheme cannot be
used in the case where the link drops the packets instead
of marking them because the throughput would be very
bad due to aggressive dropping. As in [5], we will use
in all our simulations using the GKVQ.
Experiment 2: In this experiment, we compare the perfor-
mance of the various AQM schemes assuming that the link
“marks” packets and in the presence of long-lived FTP flows
only. The queue size at the link is set to 100 packets. The
desired queue length for the REM scheme and the PI scheme
is set at 50 packets and the minthresh and the maxthresh for
the RED (with gentle turned on) scheme are set at 37 and 75
packets respectively. Recall that the desired utilization of the
link is set to be 0.98 for the AVQ scheme.
Since we use an average queue length of 50 packets for
REM and the PI controller, it is natural to attempt to regulate
the queue length to 50 for the AVQ scheme also. However, the
AVQ does not directly attempt to control queue size. Thus,
for the AVQ scheme, we drop every packet that arrives when
there are already 50 packets in the real queue. Note that this is
the worst-case scenario for the AVQ scheme, since when ECN
marking is used, the natural primary measure of performance Fig. 7. Experiment 2. Achieved utilization at the link for the different AQM
is packet loss. schemes.
We summarize our simulation results in the following.
• Packet losses and link utilization: The losses incurred that when increasing the utilization of RED (by tuning its
by all the schemes are shown in Fig. 6 as a function of parameters), the packet losses at the link also increases.
the number of FTP flows. The AVQ scheme has fewer REM and PI have a utilization of 1.0 as the queue is
losses than any other scheme except the GKVQ even always nonempty. For the AVQ scheme, we required a
at high loads. The loss rate for GKVQ and AVQ are desired utilization of 0.98 and we can see that the AVQ
comparable; however, the GKVQ marks packets more scheme tracks the desired utilization quite well. Thus, the
aggressively than any other scheme and thus has lower main conclusion from this experiment is that the AVQ
utilization. Fig. 7 shows the utilization of the link for all achieves low loss with high utilization.
the AQM schemes. Note that, the utilization of GKVQ is • Responsiveness to changing network conditions: The ob-
as low as 75%. This can once again be attributed to the jective of this experiment is to measure the response of
aggressive marking strategy of GKVQ. RED also results each AQM scheme when the number of flows is increased.
in a poor utilization of the link. Our observation has been Twenty new FTP users are added every 100 s until the
292 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 12, NO. 2, APRIL 2004

Fig. 9. Experiment 3. Evolution of the queue length at varying loads for PI.

Fig. 8. Experiment 2. Queue length at the link for varying number of FTP
connections for the different AQM schemes.

total number of FTP connections reach 180 and the av-

erage queue length over every 100 s is computed. Schemes
that have a long transient period will have an increasing
average queue length as new users are added before the
scheme is able to converge. The average queue length
(over each 100-s interval) of each scheme as the number
of users increase is shown in Fig. 8. We see from the figure
that PI and REM have higher average queue lengths than
the desired queue length. On the other hand, AVQ, GKVQ, Fig. 10. Experiment 3. Evolution of the queue sizes at varying loads for AVQ.
and RED have smaller queue sizes. This is due to the fact
that REM and PI apparently have a long transient period
and new users are added before the queue length con-
verges. The average queue length over each 100-s interval
is used to capture persistent transients in this experiment
for studying the responsiveness of the AQM schemes to
load changes. This experiment shows that AVQ is respon-
sive to changes in network load and is able to maintain
a small queue length even when the network load keeps
increasing.
Experiment 3: In this experiment, we compare the respon-
siveness of the AQM schemes when flows are dropped and then
introduced later on. Specifically, we only compare REM and
the PI controller (since these are only ones among those that
we have discussed that attempt to precisely regulate the queue
length to a desired value) with the AVQ controller. Unless oth-
erwise stated, all the system parameters are identical to Experi- Fig. 11. Experiment 3. Evolution of the queue sizes at varying loads for REM.
ment 2. The number of FTP connections is 140 at time 0.0. At
time 100, 105 FTP connections are dropped and at time 150 a time , and to the addition of flows at time 150. Fig. 11
new set of 105 FTP connections is established. We plot the evo- gives the evolution of the queue sizes for REM. The desired
lution of the queue size for each of the AQM scheme. Fig. 9 queue level in the REM scheme is 50 packets and REM is very
shows the evolution of the queue size for PI as the flows depart slow to bring the queue level to 50 packets. On removing flows,
and arrive. Note that the desired queue length is 50 packets. We the queue level drops, but on addition of new flows, there is a
can see that the system takes some time to respond to depar- large overshoot in REM.
tures and new arrivals. On the other hand, the queue in the AVQ Experiment 4: Until now, we have been comparing AVQ and
scheme in Fig. 10 responds quickly to the removal of flows at all other AQM schemes in the absence of short flows. However,
KUNNIYUR AND SRIKANT: AVQ ALGORITHM FOR ACTIVE QUEUE MANAGEMENT 293

Fig. 12. Experiment 4. Packet losses at the link for various AQM schemes. Fig. 13. Experiment 4. Utilization of the link for various AQM schemes.

a large part of the connections in the Internet comprise of short

flows. As a result, it is important to study the performance of an
AQM scheme in the presence short flows. In this experiment,
we will start with 40 FTP connections that persist throughout
the length of the experiment. We also allow the AQM schemes
to converge to the optimal solution when there are only 40 FTP
connections in the network. We then introduce short flows and
study the performance of the AQM scheme as the number of
short flows increases. We start with a short-flow arrival rate of
10 per second and gradually increase it to 50 short flows per
second. Each short flow transfers 20 packets using TCP/Reno.
The RTTs of the short flows are also distributed uniformly be-
tween the five delay classes.
We again study the following performance measures.
• Packet losses and utilization: The losses incurred by all the
schemes are shown in Fig. 12. Note that AVQ has lower
drops than the RED, REM, and the PI schemes. GKVQ in-
curs no significant packet drops (and hence cannot be seen
in the figure) because of its aggressive marking scheme.
However, as in Experiment 2, the utilization of GKVQ is
poor, as seen in Fig. 13. We again see that REM and PI Fig. 14. Experiment 4. Average queue length at the link for various AQM
schemes.
have a utilization of one, while RED and GKVQ have poor
utilization. For the AVQ scheme, the utilization is actually
slightly higher than the desired utilization at high loads, Experiment 5: In this experiment, we study the sensitivity of
but this can be attributed to the load brought by the short the AVQ algorithm to the smoothing parameter and the de-
flows. From this experiment, we can see that AVQ has sired utilization parameter . The number of long TCP connec-
lower drops compared to REM, PI and RED in the pres- tions accessing the link is fixed at 180 and the simulation setup is
ence of short flows. Even though, AVQ has more drops identical to that in Experiment 1. From Theorem 2.1, any value
than GKVQ, the utilization at the link for AVQ is signifi- of will guarantee stability. In addition to the long flows,
cantly greater than the GKVQ algorithm. short flows are introduced after 100 s. In the first part of the
• Queue length: The average queue length of each scheme experiment, the smoothing parameter is varied from 0.20 to
as the rate of the incoming short-flows short connections 0.80, while the desired utilization parameter is fixed at 0.98.
are increased is shown in Fig. 14. We see that the AVQ The evolution of the queue length is shown in Fig. 15. We can
controller maintains the smallest queue length among all see that the average queue lengths remain small irrespective of
schemes as the number of short flows increases. the value of that is used. However, the transients due to the
294 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 12, NO. 2, APRIL 2004

Fig. 15. Experiment 5. Evolution of the queue length for different values of Fig. 17. Experiment 5. Number of dropped packets versus time for different
the smoothing parameter . values of .

Fig. 18. Experiment 5. Achieved utilization versus time for different values
Fig. 16. Experiment 5. Evolution of the queue length for different values of . of .

sudden load of the short flows at s is more pronounced Note that when , packet drops are primarily caused by the
when is small. Hence, even though a very small value of sudden load change at 100 s. After a transient period in which
can guarantee stability, it can make the link sluggish to changes the algorithm tries to adapt to the new load, there are no more
in network load. A theoretical analysis of the transient behavior packet drops. But the duration of the transient period seems to
is required to precisely quantify the impact of on the system increase as is increased. When , the transients caused by
performance. This could be a topic for future research. new short flows is sufficient to cause buffer overflow in certain
In the second part of the experiment, we fix the value of to instances. The achieved utilization (averaged over one second)
be equal to 0.5 and vary the value of . We use four different is shown in Fig. 18. We can see that irrespective of the value of
values of (0.90, 0.98, 0.99, and 1.0) to study the sensitivity of , AVQ tracks the desired utilization quite well. For a desired
the algorithm to . The rest of the simulation setup is identical average queue length (or loss probability), an upper bound on
to the first part. We increase the burstiness of the short flows by the value of would depend on the traffic characteristics. Given
splitting a single short flow of 20 packets into two short flows of the traffic characteristics, desired loss probability, or desired av-
10 packets each. Fig. 16 shows the evolution of the queue length erage queue length, analytically computing an appropriate value
for the different values of . We can see that smaller values of for is a topic for future research.
results in smaller queue sizes. As a result, smaller values of re- Experiment 6: Until now, we have assumed that the router
sults in smaller number of dropped packets as shown in Fig. 17. marks packets upon detecting congestion. Instead, one can drop
KUNNIYUR AND SRIKANT: AVQ ALGORITHM FOR ACTIVE QUEUE MANAGEMENT 295

Fig. 20. Experiment 6. Utilization for various AQM schemes.

Fig. 19. Experiment 6. Average queue lengths for various AQM schemes.

packets when congestion is detected. In this experiment, we use

dropping instead of marking when the links detects an incipient
congestion event.
Note that, in the case of marking, the main goal of the adap-
tive algorithm was to match the total arrival rate to the desired
utilization of the link. However, in the case of dropping, the link
only serves those packets that are admitted to the real queue. As
a result, in the case of dropping, one adapts the virtual capacity
only when a packet has been admitted to the real queue, i.e.,
only the accepted arrival rate is taken into consideration.
We compare RED, REM, and PI controller to the AVQ
scheme. We do not use GKVQ as a dropping algorithm as the
number of packets dropped on detecting congestion would be
very high and it would result in negligible throughput. The
buffer limit at the link is set to 100 packets. The users employ
TCP NewReno. All the other parameters are as in Experiment
2. However, in this case we simulate the AVQ scheme with
both and . The reason for using
earlier was to have small losses to get the most benefit from
ECN marking. Since marking is no longer used, we also study
the AVQ under full utilization.
We have 40 FTP connections traversing the link for the en-
Fig. 21. Experiment 6. Total goodput for the various AQM schemes.
tire duration of the simulation. We allow the respective AQM
schemes to converge and then introduce short flows at 100 s.
Short flows introduced are TCP-RENO sources with 20 packets formance objectives that one would use to compare different
to transmit. The rate at which short flows arrive at the link is AQM schemes when dropping is employed as a congestion noti-
slowly increased. The average queue length and the utilization fication mechanism. In practice, we would like an AQM scheme
are shown in Fig. 19 and Fig. 20, respectively. The total goodput that maintains a small average queue length with high utiliza-
is shown in Fig. 21. By goodput, we mean the number of packets tion. However, the AQM scheme should not introduce any addi-
successfully delivered by the link to the TCP receivers. In gen- tional bias in the rates toward smaller round-trip flows (TCP by
eral, this could be different from the throughput (which is the itself introduces a bias toward smaller round-trip flows and we
total number of packets processed by the link) due to TCP’s re- do not want to add it). In this experiment, we compared the av-
transmission mechanism. Note that the average queue length, erage queue length and the utilization at the link of AVQ, RED,
the goodput of each flow and fairness are the three possible per- REM, and PI.
296 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 12, NO. 2, APRIL 2004

Note: Instead of marking or dropping a packet in the real Note that, while this is not differentiable everywhere in or
queue when the virtual queue overflows, one can mark or drop , it is differentiable in the region . Substituting for
packets in the real queue by applying RED (or any other AQM , and and using the fact that
algorithm) in the virtual queue. Thus, if there are desirable fea- , we find that
tures in other AQM schemes, they can be easily incorporated in
the AVQ algorithm. When marking is employed, our experience (9)
is that a simple mark-tail would be sufficient as shown in Exper-
iments 1 through 4. In the case when the link drops the packets, Note that . Let denote the Laplace transform
many successive packet drops from the same flow could cause of and let denote the Laplace transform of .
timeouts. To avoid this, one could randomize the dropping by Taking the Laplace transforms of (4) and (5), we get
using a mechanism like RED in the virtual queue to prevent
bursts of packets of the same flow to be dropped. Hence, in this (10)
experiment, we also consider an AVQ scheme that had RED im- (11)
plemented in its virtual queue.
Our experience has been that, if RED is employed in the vir- Substituting (11) in (10), we get the so-called characteristic
tual queue, the performance of the AQM scheme is not very sen- equation
sitive to the choice of the RED parameters. Even though, there
is no significant difference in the goodputs, we believe that the (12)
fairness of the AVQ scheme can be improved using a proba-
bilistic dropping scheme in the virtual queue at the expense of Next, we use the fact that the roots are continuous functions of
increased computation at the router. We intend to study the fair- the round-trip delay . As a result, if the system is stable with
ness properties in our future research work. Note that a proba- for a fixed value of , then the roots are strictly in the
bilistic AQM scheme on the virtual queue is required only when left half-plane. Therefore, we can choose small enough such
the link drops packets and not when the link marks packets be- that the roots still remain in the left half-plane. This will help
cause multiple marks within a single window does not cause us to find the maximum feedback delay for which the system is
TCP to time out or go into slow-start. stable for a given . We will then show that we can use the same
technique to show that given a feedback delay , one can find
IV. STABILITY ANALYSIS OF THE AVQ SCHEME the maximum value of for which the system is stable. We will
formalize these ideas in this section.
In this section, we will prove the main result of the paper,
When , i.e., there is no feedback delay in the system,
which was stated in Theorem 2.1. The starting point of the anal-
the characteristic equation reduces to
ysis is the linearized version of the TCP/AQM model derived in
(4) and (5). We summarize the main ideas behind the proof as
follows. (13)
• The stability of a linear delay-differential equation can Solving the quadratic equation, we get
be analyzed using its characteristic equation. The charac-
teristic equation of the linear delay-differential equation
can be obtained by taking its Laplace transform. For the
linearized system to be stable, its characteristic equation
should have all its roots in the left half-plane (i.e., if is a If , then the system has all real roots
root of the characteristic equation, then ). which lie strictly in the left half-plane. If
• We will first show that for , , and fixed, the system is , then the system has complex roots that also lie strictly
stable in the absence of feedback delays, i.e., . This in the left half-plane. Thus, for all values of , the system
implies that all the roots of the characteristic equation lie in is stable.
the left half-plane. The roots of the characteristic equation The following theorem gives the necessary condition on the
are continuous functions of its parameters. Therefore, the RTT for the stability of the system given by (4) and (5).
roots of the characteristic equation are continuous function Theorem 4.1: Fix , the number of TCP users , and
of the feedback delay . By increasing , one can find the the utilization . Find the smallest such that
smallest feedback delay at which one of the roots hits
the imaginary axis (if there is no such , then the system is
stable for all .). Hence, for all , the system has all
its roots in the left half-plane and, hence, it is stable. This (14)
is the key idea behind the stability analysis in this section. satisfies
Recall that the linearized TCP/AQM system was given in (4)
and (5). For analytical tractability, we assume that (15)

for some . Then, the TCP/AQM system given in

(8) (4) and (5) is stable for all values of .
KUNNIYUR AND SRIKANT: AVQ ALGORITHM FOR ACTIVE QUEUE MANAGEMENT 297

Proof: The characteristic equation of the TCP/AQM Theorem 4.2: Fix , the number of TCP users , and
system (12) can be rewritten as the utilization . Define to be

(16) (19)
Let be one of the roots of the characteristic equation at the
smallest such that the roots hits the imaginary axis. where solves
Since the roots on the imaginary axis are complementary, we (20)
will concern ourselves only with . From (16), we get
and is as given in (14). Then for all the system is stable.
Moreover, is unique.
Proof: Note that and do not depend on . Also,
To satisfy the last condition, the following conditions must be . Therefore, as increases, decreases
met simultaneously. and increases. Note that we can easily show that is a unique
Condition on magnitude: solution to (20). Also, note that . Hence

Let solve (15) for some . Then, we claim that

Condition on angles: (21)
Suppose not. Since and , we have
. Thus, using Fact 4.1
From the condition on magnitude, we get

Therefore

(22)
(or)
Also, since , . Thus

(17) But this contradicts (22). Hence, . Thus, any also

From the condition on angles, we get satisfies , and therefore, for any , the system is stable
from Theorem 4.1.
Until now, we have been given a fixed and a fixed and
we were interested in finding the largest feedback delay for
for . Since is a decreasing function of , and
which this system is stable. But, a more practical question is
and are independent of , we note that
the following: given a feedback delay , and number of users
is an increasing function of . Therefore, the smallest
, how can one design such that the system is stable? The
that solves (15) gives the smallest delay such that at least one of
next theorem gives a method by which one can design such
the roots hits the imaginary axis. Therefore, for all , the
that system is stable. Note that this theorem is the main result
system is locally asymptotically stable.
of the paper and is also stated in Section II. We state it again for
Remark: Although Theorem 4.1 provides a necessary and
convenience.
sufficient condition, it is hard to verify the conditions of the
Theorem 4.3: Suppose that the feedback delay , number of
theorem numerically due to the following issues:
users , and the utilization , are given. Let be given by
• What value of will yield the smallest ?
• If solves (15), how can we be sure that there exists no ,
such that solves (15) and ? (23)
Theorem 4.2 solves these issues by giving an easily verifiable
sufficient condition for stability. Before stating the theorem, we Then, for all , the system is locally stable.
state the following useful fact. Proof: Using (20), we know that
Fact 4.1: Let and be arbitrary positive constants with
. Then
(18) Now let us fix an . Since, , we have
298 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 12, NO. 2, APRIL 2004

and . Using (23) to solve for , we get . Now,

if we increase the delay to 0.25 s, there exists no that will solve
(24). However, using (23) to solve for , we get .
Therefore, using Fact 4.1, we get for all In view of this example and the previous theorem, (23) is a less
restrictive condition on than (24).
The next theorem quantifies the impact of , the number of
users, on the stability of the system.
Theorem 4.6: Fix the feedback delay , the smoothing pa-
rameter and the utilization . Define to be
(26)
Hence, the system is stable for all .
In [12], we presented the following condition for stability. where satisfies
Theorem 4.4: Fix the feedback delay , the number of users
, and the utilization . Find satisfying

(24) and is as given in (14). Then, for all , the system is

stable.
where is as given in (14). Then, for all , the system is Proof: Note that in this case, , , and are all
stable. functions of . We can easily show that as increases,
We will now show that if there exists an that satisfies increases, decreases, decreases and de-
(24), then there exists an that satisfies (23). creases. Using this and following along the lines of the proof
Theorem 4.5: Fix the feedback delay , the number of users for Theorem 4.3, we can show that for all , the system
, and the utilization . Let solve (24). Then there exists is locally stable.
an that satisfies (23). A similar theorem can now be stated for .
Proof: Suppose there does not exist an that satis- Theorem 4.7: Fix the feedback delay , number of users
fies (23). This implies and the smoothing parameter . Define to be

(25) (27)

However, there exists an such that where solves

Let and is as given in (14). Then, for all , the system is

stable.

V. CONCLUSION

Therefore Robust Active Queue Management schemes at the routers are

essential for a low-loss, low-delay network. In this paper, an
easily implementable robust AQM scheme called the Adaptive
Virtual Queue (AVQ) algorithm is presented. The implementa-
From (17), we know that tion complexity of the AVQ algorithm is comparable to other
well-known AQM schemes. While the present form of AVQ
adapts the virtual capacity on every packet arrival, an actual im-
Therefore plementation in routers might perform this adaptation every
packets (as in the PI controller [7]). Parameter choice for AVQ
in this scenario is a topic for future research.
In this paper, we provide simple design rules to choose the
This contradicts (25). Hence, there exists a that satisfies smoothing parameters of the AVQ algorithm. The criterion we
(23). use to choose the parameters is local stability of the congestion
Example 4.1: In this example, we will compare the value controllers and the AQM scheme together. However, it would
of obtained from (24) and (23). In particular, we will show also be useful to quantify the impact of the desired utilization
that one might not be able to solve (24) even though one can on QoS parameters such as loss probability or average queue
solve (23). Consider a single link with 10-Mb/s capacity and length. This is a topic for future research.
let there be 50 users accessing it. Using an average packet size We show through simulations that the AVQ controller out per-
of 1000 bytes, the capacity of the link can be written as 1250 forms a number of other well-known AQM schemes in terms of
packets per second. Let the desired utilization at the link to losses, utilization and average queue length. In particular, we
be 0.98. Let . Using (24) we compute the value of show that AVQ is able to maintain a small average queue length
KUNNIYUR AND SRIKANT: AVQ ALGORITHM FOR ACTIVE QUEUE MANAGEMENT 299

at high utilizations with minimal loss at the routers. This con- [12] , “Analysis and design of an adaptive virtual queue (AVQ) algo-
clusion also holds in the presence of short flows arriving and de- rithm for active queue management,” in Proc. ACM SIGCOMM, San
Diego, CA, Aug. 2001.
parting at the link. We also show that AVQ responds to changing [13] , “A time-scale decomposition approach to adaptive ECN
network conditions better than other AQM schemes (in terms of marking,” in Proc. IEEE INFOCOM, Anchorage, AK, Apr. 2001, pp.
average queue length, utilization, and losses). 1330–1339.
[14] L. Massoulie and J. Roberts, “Bandwidth sharing: objectives and
We also study the performance of AVQ when dropping (in- algorithms,” in Proc. IEEE INFOCOM, New York, Mar. 1999, pp.
stead of marking) is employed at the routers. While AVQ per- 1395–1403.
forms better than other AQM schemes in terms of utilization [15] V. Misra, W. Gong, and D. Towlsey, “A fluid-based analysis of a network
of AQM routers supporting TCP flows with an application to RED,” in
and average queue length, the fairness of AVQ can be improved Proc. ACM SIGCOMM, Stockholm, Sweden, Sept. 2000, pp. 151–160.
using a probabilistic AQM scheme (like RED) on AVQ. We note [16] T. J. Ott, T. V. Lakshman, and L. H. Wong, “SRED: stabilized RED,” in
that a probabilistic AQM scheme on the virtual queue is required Proc. IEEE INFOCOM, New York, Mar. 1999, pp. 1346–1355.
[17] J. Padhye, V. Firoiu, D. Towsley, and J. Kurose, “Modeling TCP
only when the link drops packets and not when the link marks throughput: a simple model and its empirical validation,” in Proc. ACM
packets because multiple marks within a single window does not SIGCOMM, Vancouver, BC, Canada, 1998, pp. 303–314.
cause TCP to time out or go into slow-start. The fairness prop- [18] S. Shakkottai and R. Srikant, “Mean FDE models for Internet congestion
control under a many-flows regime,” IEEE Trans. Inform. Theory, 2004,
erties of AVQ with packet dropping is a topic of future research. to be published.
An important feature of the AVQ algorithm is that one can [19] S. Shakkottai and R. Srikant, “How good are deterministic fluid models
employ any AQM algorithm in the virtual queue. Thus, if there of Internet congestion control?,” in Proc. IEEE INFOCOM, 2002, pp.
497–505.
are desirable properties in any other marking schemes, one can [20] S. Floyd. (1997) RED: Discussions of setting parameters. [Online].
easily incorporate it into the AVQ scheme. However, when Available: [Link]
marking is employed, our experience has been that a simple
mark-tail would suffice.
Srisankar S. Kunniyur (M’98) received the B.E. de-
REFERENCES gree in electrical and electronics engineering from
B.I.T.S., Pilani, India, in 1996 and the M.S. and Ph.D.
[1] S. Athuraliya, D. E. Lapsley, and S. H. Low, “Random early marking degrees in electrical engineering from the University
for Internet congestion control,” in Proc. IEEE Globecom, 1999, pp.
of Illinois at Urbana-Champaign in 1998 and 2001,
1747–1752. respectively.
[2] W. Feng, D. Kandlur, D. Saha, and K. Shin, “Blue: A New Class of
He is currently an Assistant Professor of electrical
Active Queue Management Schemes,” Univ. Michigan, Tech. Rep. and systems engineering at the University of Penn-
CSE-TR-387-99, Apr. 1999.
sylvania, Philadelphia. His research interests include
[3] S. Floyd, “TCP and explicit congestion notification,” ACM Comput. design and performance analysis of communication
Commun. Rev., vol. 24, pp. 10–23, Oct. 1994.
networks, congestion control, and pricing in hetero-
[4] S. Floyd and V. Jacobson, “Random early detection gateways for con- geneous networks.
gestion avoidance,” IEEE/ACM Trans. Networking, vol. 1, pp. 397–413,
Aug. 1993.
[5] R. J. Gibbens and F. P. Kelly, “Distributed connection acceptance con-
trol for a connectionless network,” in Proc. 16th Int. Teletraffic Congr.,
Edinburgh, Scotland, June 1999, pp. 941–952. R. Srikant (M’91–SM’01) received the [Link]. de-
[6] C. V. Hollot, V. Misra, D. Towlsey, and W. Gong, “A control theo- gree from the Indian Institute of Technology, Madras,
retic analysis of RED,” in Proc. IEEE INFOCOM, Anchorage, AK, Apr. in 1985, and the M.S. and Ph.D. degrees from the
2001, pp. 1510–1519. University of Illinois at Urbana-Champaign in 1988
[7] , “On designing improved controllers for AQM routers supporting and 1991, respectively, all in electrical engineering.
TCP flows,” in Proc. IEEE INFOCOM, Anchorage, AK, Apr. 2001, pp. He was a Member of Technical Staff at AT&T
1726–1734. Bell Laboratories from 1991 to 1995. He is currently
[8] F. P. Kelly, “Mathematical modeling of the Internet,” in Mathe- with the University of Illinois at Urbana-Champaign
matics Unlimited—2001 and Beyond, B. Engquist and W. Schmid, where he is an Associate Professor in the Depart-
Eds. Berlin, Germany: Springer-Verlag, 2001, pp. 685–702. ment of Electrical and Computer Engineering and
[9] F. P. Kelly, P. Key, and S. Zachary, “Distributed admission control,” a Research Associate Professor in the Coordinated
IEEE J. Select. Areas Commun., vol. 18, pp. 2617–2628, Dec. 2000. Science Laboratory. He was an associate editor of Automatica. His research
[10] S. Kunniyur and R. Srikant, “End-to-end congestion control: utility interests include communication networks, stochastic processes, queueing
functions, random losses and ECN marks,” in Proc. IEEE INFOCOM, theory, information theory, and game theory.
Tel Aviv, Israel, Mar. 2000, pp. 1323–1332. Dr. Srikant is currently on the editorial boards of the IEEE/ACM
[11] S. Kunniyur and R. Srikant, “End-to-end congestion control: utility TRANSACTIONS ON NETWORKING and IEEE TRANSACTIONS ON AUTOMATIC
functions, random losses and ECN marks,” IEEE/ACM Trans. Net- CONTROL. He was the Chair of the 2002 IEEE Computer Communications
working, vol. 11, pp. 689–702, Oct. 2003. Workshop in Santa Fe, NM.