Nonlinear RED: A simple yet ecient active queue

management scheme
q
Kaiyu Zhou
*
, Kwan L. Yeung, Victor O.K. Li
Department of Electrical and Electronic Engineering, The University of Hong Kong, CYC805, HKU,
Pokfulam Road, Hong Kong, China
Received 1 June 2005; received in revised form 26 January 2006; accepted 13 April 2006
Available online 17 May 2006
Responsible Editor: Chin-T. Lea
Abstract
Among various active queue management schemes (AQM), random early detection (RED) is probably the most exten-
sively studied. Unlike the existing RED enhancement schemes, we replace the linear packet dropping function in RED by a
judicially designed nonlinear quadratic function. The rest of the original RED remains unchanged. We call this new
scheme Nonlinear RED, or NLRED. The underlying idea is that, with the proposed nonlinear packet dropping function,
packet dropping becomes gentler than RED at light trac load but more aggressive at heavy load. As a result, at light
trac load, NLRED encourages the router to operate in a range of average queue sizes rather than a xed one. When
the load is heavy and the average queue size approaches the pre-determined maximum threshold (i.e. the queue size
may soon get out of control), NLRED allows more aggressive packet dropping to back o from it. Simulations demon-
strate that NLRED achieves a higher and more stable throughput than RED and REM, another ecient variant of RED.
Since NLRED is fully compatible with RED, we can easily upgrade/replace the existing RED implementations by
NLRED.
2006 Elsevier B.V. All rights reserved.
Keywords: Active queue management; Nonlinear RED; Random early marking
1. Introduction
Congestion control is one of the most important
problems in the Internet. Most of the existing Inter-
net routers play a passive role in congestion control,
and are known as drop-tail routers. A drop-tail rou-
ter discards packets when its FIFO queue is full. It
was shown in [1] that under heavy load conditions,
drop-tail routers cause global synchronization, a
phenomenon in which all senders sharing the same
1389-1286/$ - see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.comnet.2006.04.007
q
This research is supported in part by the Areas of Excellence
Scheme established under the University Grants Committee of
the Hong Kong Special Administrative Region, China (Project
No. AoE/E-01/99), and the Research Grant Council Earmarked
Grant 7048/02E.
*
Corresponding author. Tel.: +852 28592692.
E-mail addresses: kyzhou@eee.hku.hk (K. Zhou), kyeung@
eee.hku.hk (K.L. Yeung), vli@eee.hku.hk (V.O.K. Li).
Computer Networks 50 (2006) 37843794
www.elsevier.com/locate/comnet
bottleneck router/link shut down their transmission
windows at almost the same time, thereby causing a
sharp drop in the bottleneck link utilization. It was
also found in [2] that drop-tail routers are biased
against bursty sources. This is because, when a burst
of packets from a sender arrives at a fully occupied
queue, a sustained packet drop within the same
window of data occurs. Research in [3,4] showed
that the dominant transport layer protocol, TCP
[5], lacks the ability to recover from such multiple
packet losses within the same window of data.
Therefore, the TCP sender has to rely on retrans-
mission timeouts to recover the lost packets.
Retransmission timeout signicantly slows down
the transmission rate of a TCP ow, because almost
no data will be sent when the sender waits for the
retransmission timer to expire. A good congestion
control scheme should therefore avoid triggering
unnecessary timeouts.
On the other hand, as more and more multimedia
applications running on top of UDP are being
deployed, the traditional approach of relying solely
on TCPs end-to-end congestion control algorithms
will no longer be viable [6]. The network, in partic-
ular the routers in the network, should play an
active role in its resource allocation, so as to eec-
tively control/prevent congestion. This is known as
active queue management (AQM) [7]. The essence
is that an AQM router may intelligently drop pack-
ets before the queue overows.
Among various AQM schemes, random early
detection (RED) [8] is probably the most exten-
sively studied. RED is shown to eectively tackle
both the global synchronization problem and the
problem of bias against bursty sources. Due to its
popularity, RED (or its variants) has been imple-
mented by many router vendors in their products
(e.g. Cisco implemented WRED [9]). On the other
hand, there is still a hot on-going debate on the
performance of RED. Some researchers claimed
that RED appears to provide no clear advantage
over drop-tail mechanism [10]. But more research-
ers [2,1124] acknowledged that RED shows some
advantages over drop-tail routers but it is not per-
fect, mainly due to one or more of the following
problems.
RED performance is highly sensitive to its para-
meter settings [11,12,17,18,21,23,25]. In RED
(as detailed in Section 2), at least 4 parameters,
namely, maximum threshold (max
th
), minimum
threshold (min
th
), maximum packet dropping
probability (max
p
), and weighting factor (x
q
),
have to be properly set.
RED performance is sensitive to the number of
competing sources/ows [12,1719,21,25].
RED performance is sensitive to the packet size
[15].
With RED, wild queue oscillation is observed
when the trac load changes [12,22].
As a result, RED has been extended and
enhanced in many dierent ways [2,1121]. It can
be found that a common underlying technique
adopted in most studies is to steer a router to oper-
ate around a xed target queue size (which can
either be an average queue size or an instantaneous
queue size). There are some concerns on the suit-
ability of this approach, since the schemes thus
designed are usually more complicated than the
original RED. This renders them unsuitable for
backbone routers where ecient implementation is
of primary concern. In some schemes (e.g. [13]),
additional parameters are also introduced. This
adds extra complexity to the task of parameter
setting.
Unlike the existing RED enhancement schemes,
we propose to simply replace the linear packet drop-
ping function in RED by a judiciously designed
nonlinear quadratic function. The rest of the origi-
nal RED remains unchanged. We call this new
scheme Nonlinear RED, or NLRED. The underly-
ing idea is that, with the proposed nonlinear packet
dropping function, packet dropping is gentler than
RED at light trac load but more aggressive at
heavy load. Therefore, at light trac load NLRED
encourages the router to operate in a range of aver-
age queue sizes rather than a xed one. When the
load is heavy and the average queue size approaches
the maximum threshold max
th
an indicator that
the queue size may soon get out of control, NLRED
allows more aggressive packet dropping to quickly
back o from it. Simulations demonstrate that
NLRED achieves a higher and more stable through-
put than RED and REM [16], an ecient variant of
RED. Since NLRED is fully compatible with RED,
we can easily upgrade/replace the existing RED
implementations by NLRED.
In the next section, some major RED enhance-
ment schemes are reviewed, which provides the nec-
essary background for NLRED design in Section 3.
In Section 4, NLRED performance is studied using
simulations. Finally, we conclude the paper in
Section 5.
K. Zhou et al. / Computer Networks 50 (2006) 37843794 3785
2. Related work
RED [8] was mainly designed to overcome the
two problems associated with drop-tail routers,
namely, global synchronization and bias against
bursty sources. Unlike the drop-tail mechanism,
RED measures congestion by the average queue size
and drops packets randomly before the router queue
overows. When a packet arrives at a router, the
average queue size, denoted avg, is updated using
the following exponentially weighted moving aver-
age (EWMA) function:
avg 1 x
q
avg
0
x
q
q;
where avg
0
is the calculated average queue size when
the last packet arrived, q is the instantaneous queue
size, and x
q
is the pre-determined weighting factor
with a value between 0 and 1.
As avg varies from a minimum threshold min
th
to
a maximum threshold max
th
, the packet dropping
probability p
d
increases linearly from 0 to a maxi-
mum packet dropping probability max
p
, or
p
d

0 avg < min
th
;
avgmin
th
max
th
min
th
max
p
min
th
6 avg < max
th
;
1 max
th
6 avg:
_

_
1
May et al. [13] modeled the throughput of RED and
they showed that under heavy load the throughput
is inversely proportional to the load. This problem
can be solved by selecting appropriate parameters
for RED [11]. Besides, queuing delay and jitter have
also been studied. It is found that delay jitter in
RED is very sensitive to the value of the weighting
factor x
q
. Small x
q
results in large delay variance,
which is not suitable for real-time applications [22].
Zheng and Atiquzzaman [14] believed that the
most important problem of RED is low throughput.
They found that just modifying the parameters of
RED for throughput enhancement is not sucient;
so they proposed a scheme that uses a packet drop-
ping function consisting of two linear segments
instead of a single one. This change is in fact very
similar to that of gentle RED (GRED) by Floyd
[26]. In the original RED, the packet dropping
probability is set to 1 when max
th
6 avg. This was
later considered to be too aggressive and thus a gen-
tle version of RED (or GRED) was proposed,
which is implemented in the ns-2 [27]. In GRED,
the packet dropping probability increases linearly
from max
p
to 1 when avg increases from max
th
to
2max
th
. The resulting packet dropping function is
shown in Fig. 1.
In [18], Feng et al. showed that the eectiveness
of dropping packets to control source rate is aected
by the number of sources. So there is no single set of
RED parameters that can work well under dierent
congestion scenarios. Therefore, they proposed the
adaptive RED (ARED) scheme, which adapts the
max
p
value based on the trac load.
In [17], Floyd acknowledged that the average
queue size of the original RED is sensitive to the
level of congestion as well as the RED parameter
settings. She then introduced a new adaptive RED
scheme, which is partially based on ARED. In this
scheme, the variance of average queue size is used
to adjust max
p
. It also diers from ARED in that
the average queue size is steered to a target around
(max
th
+ min
th
)/2. Besides, instead of using a con-
stant weighting factor x
q
, x
q
is given by x
q
= 1
e
1/L
, where L is the capacity of the bottleneck link
under consideration.
In [12], Teunis J. Ott et al. observed wild oscilla-
tions in instantaneous queue size when the trac
load changes. The trac load is measured by the
number of active ows carried by a link. Therefore,
they designed a mechanism to estimate the number
of active ows in a bottleneck link. Based on the
estimated number of ows, an adaptive AQM
scheme was proposed, aiming at operating the rou-
ter at a xed target queue size.
In [16], by arguing that the average queue size is a
performance measure instead of a congestion mea-
sure, an approach known as random exponential
marking (REM) was developed and analyzed. It
was shown that REM is able to achieve high link
utilization, negligible packet loss, and short queuing
delay in a simple and scalable manner. In REM, a
state variable r is maintained via
Fig. 1. Dropping functions for RED and GRED.
3786 K. Zhou et al. / Computer Networks 50 (2006) 37843794
rk 1 rk caqk b kk C T;
2
where b is the desired queue occupancy, k(k) is the
arrival rate during the kth sample, T is the sample
period, C is the link capacity, c and a are constants.
The mapping from r to the packet drop probability
is
p
rem
k 1 /
rk
; 3
where / is a constant parameter.
In [20], AVQ takes a rather dierent approach to
design an AQM. AVQ seeks to keep the queue
empty. To this end, a virtual link speed is intro-
duced, which is kept no greater than the actual link
speed. Accordingly, a virtual queue is maintained
along with the virtual link speed. As packets arrive,
they are placed in the real queue while a token is
placed in the virtual queue. The real packets are
served with the link speed, while the tokens are
served at the virtual link speed. The service time
of a token is the size of the corresponding packet
divided by the virtual link speed. If a token nds
the virtual queue full, then the real packet and token
are dropped. The virtual capacity at each link is
then adapted to ensure that the total ow entering
each link achieves a desired utilization of the link.
Modeling the dynamics of TCP and AQM is
another important research area. In particular,
Misra et al. modeled the dynamics of TCP and
AQM with stochastic dierential equations in [15].
They showed that it is the combination of link
bandwidth, average packet size, sampling interval,
and trac load that make RED stable. Partly based
on those results, in [19], Hollot et al. used control
theory to investigate AQM as a classical feedback
system and introduced the PI controller to improve
REDs performance. Recently, Gao and Hou [23]
argued that the model of TCP and AQM dynamics
in the previous work does not consider the real
implementation of a TCP; so they proposed a state
feedback controller (SFC).
To summarize, we note that schemes in [12,15
17,20,21,23] tend to be more complicated than the
original RED. They all adopted a common underly-
ing technique to steer a router to operate at a xed
target queue size independent of trac load. We
have some concerns on the suitability of this
approach. Suppose the target queue size is set too
small and the oered load is high (so the number
of ows is large), each active ows congestion win-
dow size will be very small due to resource sharing.
As a result, when a ow has less than three packets
acknowledged in a single window, the ow has to
wait for retransmission timeout, and the throughput
is low. On the other hand, if the target queue size is
set too large and the oered load is light, packets
will experience a long queuing delay without any
gain in throughput. If the target queue size is
allowed to vary with the trac load, we observe
the following advantages:
The potential problems with the xed target
queue size can be removed.
A fairer congestion control can be realized.
To illustrate the second point, we can express the
throughput of a TCP connection as
R
w MSS
RTT
; 4
where w is the sender window size, MSS is the seg-
ment size, and RTT is the round-trip-time (which
increases with the queue size). From (4), increasing
RTT reduces the throughput, and thus the sending
rate. Therefore, allowing router queue size to build
up is another way, probably with a ner granularity
of control, to slow down all senders passing through
this router when congestion occurs. Therefore, this
tends to be fairer than just punishing a few unlucky
senders by random packet dropping.
3. Nonlinear random early detection
3.1. NLRED Algorithm
The throughput performance of RED is not sta-
ble. For example, when the trac load is very light
and RED parameters are aggressively set or when
the trac load is very heavy and the parameters
are tenderly set, the throughput is low. It has been
shown that no single set of parameters for RED
could get a stable performance under dierent trac
loads. We believe such instability is due, at least in
part, to the linear packet dropping function adopted
by RED, which tends to be too aggressive at light
load, and not aggressive enough when the average
queue size approaches the maximum threshold
max
th
. We also believe that the performance im-
provement of some previous work is at least partly
due to the employment of nonlinear dropping func-
tion (e.g. [14,16]), either intentionally or uninten-
tionally. (More reasons to be provided later.)
However, we notice that these improvements may
K. Zhou et al. / Computer Networks 50 (2006) 37843794 3787
not be suitable for core routers, as their nonlinear
dropping functions greatly complicate the basic
mechanism of RED. In this paper, we propose to
replace the linear packet dropping function by a
judicially designed quadratic function. The resulting
scheme is called non-linear RED or NLRED. The
pseudocode of NLRED is summarized in Fig. 2.
When avg exceeds the minimum threshold,
NLRED uses the nonlinear quadratic function
shown in (5) to drop packets, where max
0
p
represents
the maximum packet dropping probability of
NLRED. Fig. 3 compares the packet dropping
functions for RED and NLRED. (The choice of a
quadratic function is further explained in the next
subsection.)
p
0
d

0 avg 6 min
th
;
avgmin
th
max
th
min
th
_ _
2
max
0
p
min
th
< avg 6 max
th
;
1 max
th
< avg:
_

_
5
Comparing (5) to the dropping function of original
RED in (1), if the same value of max
p
is used,
NLRED will be gentler than RED for all trac
load. This is because the packet dropping probabil-
ity of NLRED will always be smaller than that of
RED. In order to make the two schemes to have a
comparable total packet dropping probabilities, we
set max
0
p
1:5max
p
, such that the areas covered
by both dropping functions from min
th
to max
th
are the same, or
_
max
th
min
th
p
d
davg
_
max
th
min
th
p
0
d
davg:
3.2. Why use a quadratic function?
Given that N TCP ows equally share a link with
bandwidth L, and experience a random packet loss/
drop probability p. It was shown [18] that p and N
has the following relationship:
p <
N
L
MSS a
RTT
_ _
2
;
where a is a constant. This equation indicates that
to eectively manage the ows (so as to fully utilize
the available network bandwidth) the packet drop-
ping probability should vary quadratically with
the number of ows. However, nding the number
of active ows N needs (i) per ow information,
(ii) extra storage space for storing extra state infor-
mation, and (iii) extra router processing overhead.
Besides, the resulting ow number is nothing more
than an estimation [12,24].
In (5), we have proposed to vary the packet drop-
ping probability based on a quadratic function of
average queue size. In [19], it is shown that the
average queue size at a router is roughly directly
NLRED
for each packet arrival:
calculate the average queue size avg
if
th
avg min
no packet drop
else if
th th
min avg max
calculate the packet drop probability using (3)
drop the packet with the calculated probability
else
drop the packet
<
< <
Fig. 2. Pseudocode of NLRED.
Fig. 3. Dropping functions for NLRED and RED.
60
65
70
75
80
85
90
95
0 50 100 150
Flows
A
v
e
r
a
g
e

q
u
e
u
e

l
e
n
g
t
h

(
p
a
c
k
e
t
s
)
Fig. 4. Average queue size vs. number of ows, with drop-tail
router.
3788 K. Zhou et al. / Computer Networks 50 (2006) 37843794
proportional to the number of active TCP ows
passing through it. This is further veried by the
simulations results shown in Fig. 4. The average
queue size versus the number of ows is obtained
by simulating the network in Fig. 5 with drop-tail
router mechanisms. (Other simulations using RED
with dierent trac load also show similar results.)
In fact, choosing a quadratic function is also
intuitively appealing. From Fig. 3, when the average
queue size is slightly larger than min
th
, the packet
dropping probability is smaller than the corre-
sponding RED. As such, the average queue size will
not be forced to work around min
th
as strongly as
that in RED. Or, one can interpret this as follows.
Under current trac load, the signal for congestion
is not strong enough to justify any severe measures
to cut back queue size; so a gentler than RED
packet dropping probability is desirable. While
doing this, we naturally encourage the routers to
operate over a range of queue sizes closer to min
th
(instead of at a xed target queue size). When avg
approaches max
th
, the congestion becomes more
pronounced. The routers can thus take decisive
actions to drop packets at a rate higher than RED.
4. Simulation results
NLRED is implemented using ns-2 simulator
[27]. We conduct the simulations based on the net-
work in Fig. 5, which consists of N senders and
one sink, connected together via two routers A
and B. The link between the two routers is the bot-
tleneck. By varying N, we produce dierent levels of
trac load and thus dierent levels of congestion on
the bottleneck link. The active queue management
schemes under investigation are implemented at
router A, which has a queue buer size of 120 pack-
ets. Unless otherwise stated, we assume that all
packets generated by the senders are 1000 bytes
long. Extensive simulations based on this network
using dierent TCP implementations (Tahoe, Reno,
and New Reno), RTTs, and AQM schemes (with
dierent parameter sets), are conducted, whereas
only a representative subset of the results based on
TCP Reno is reported below. Besides, we choose
to compare NLRED with GRED [26] instead of
RED, due to the superior performance of GRED
over RED. We also compare NLRED with REM
[27] as it is a representative scheme that steers a rou-
ter to operate around a xed target queue size and
reported with excellent performance.
4.1. Experiment 1
Figs. 69 show the results of a set of simulations
with the number of long-lived TCP ows increasing
from 5 to 120 and max
p
varying from 0.02 to 0.5.
The receivers advertised window of each connec-
tion is set to be bigger than the bandwidth delay
product. Each point of the simulation results is
Fig. 5. The network simulated.
T
h
r
o
u
g
h
p
u
t
/
B
a
n
d
w
i
d
t
h

(
%
)
0 20 40 60 80 100 120
70
75
80
85
90
95
100
Flows
max
p
=0.5
max
p
=0.2
max
p
=0.125
max
p
=0.1
max
p
=0.05
max
p
=0.03
max
p
=0.025
max
p
=0.02
Fig. 6. Throughput vs. number of ows using GRED.
K. Zhou et al. / Computer Networks 50 (2006) 37843794 3789
obtained from a single 200 s simulation while the
statistics are collected in the second half of the sim-
ulation time (i.e. the second 100-s interval).
As explained in Section 3.1, in order to compare
GRED and NLRED, the maximum packet drop-
ping probability of NLRED is set as max
0
p

1:5max
p
. As such, the simulation results/curves
obtained using NLRED will be labelled by its equiv-
alent max
p
instead of max
0
p
. As an example, the line
labelled with max
p
= 0.1 in Fig. 9 means the actual
maximum packet dropping probability is
max
0
p
0:15. Both GRED and NLRED use the
same set of parameters, x
q
= 0.002, min
th
= 10,
and max
th
= 30.
Figs. 6 and 7 show the bottleneck link throughput
against the number of ows. Each curve in the g-
ures represents the simulation results with a given
max
p
. Comparing the two gures, we can see that
NLRED is less sensitive to the choice of max
p
under
dierent trac loads (i.e. number of ows).
Although the throughput of NLRED still changes
with the load, for some max
p
selections (e.g.
max
p
= 0.050.1, or max
0
p
0:0750:15), NLRED
is very successful in maintaining a high throughput
regardless of the loading. This is mainly due to
NLREDs nonlinear quadratic packet dropping
function, which allows more packet bursts to pass
when the average queue size is small, and drops more
packets when the average queue size becomes large.
Figs. 8 and 9 show the change of the average
queue size with the number of ows. Unlike GRED,
we can see that NLRED allows the average queue
size to grow at a faster rate when the number of
ows is small. As the number of ows increases,
NLRED tends to control the average queue size bet-
ter (i.e. the queue size converges to a stable value
faster) than GRED.
To have a closer examination on the ability to
control queue size, we show in Figs. 10 and 11 the
instantaneous and average queue sizes against time,
with the number of ows N = 100 and max
p
= 0.02.
We can see that the oscillations in both instanta-
neous and average queue sizes are much more
noticeable when GRED is used. With NLRED,
the oscillations are eectively suppressed, again
due to its nonlinear packet dropping function.
4.2. Experiment 2
We compare the performance of GRED, REM
[28] and NLRED under dierent trac loads. We
set max
p
of all the three AQM schemes to 0.1,
Flows
T
h
r
o
u
g
h
p
u
t
/
B
a
n
d
w
i
d
t
h

(
%
)
0 20 40 60 80 100 120
75
80
85
90
95
100
max
p
=0.5
max
p
=0.2
max
p
=0.125
max
p
=0.1
max
p
=0.05
max
p
=0.03
max
p
=0.025
max
p
=0.02
Fig. 7. Throughput vs. number of ows using NLRED.
0 20 40 60 80 100 120
5
10
15
20
25
30
35
40
Flows
A
v
e
r
a
g
e

q
u
e
u
e

l
e
n
g
t
h

(
p
a
c
k
e
t
s
)
max
p
=0.5
max
p
=0.2
max
p
=0.125
max
p
=0.1
max
p
=0.05
max
p
=0.03
max
p
=0.025
max
p
=0.02
Fig. 8. Average queue size vs. number of ows using GRED.
A
v
e
r
a
g
e

q
u
e
u
e

l
e
n
g
t
h

(
p
a
c
k
e
t
s
)
0 20 40 60 80 100 120
5
10
15
20
25
30
Flows
max
p
=0.5
max
p
=0.2
max
p
=0.125
max
p
=0.1
max
p
=0.05
max
p
=0.03
max
p
=0.025
max
p
=0.02
Fig. 9. Average queue size vs. number of ows using NLRED.
3790 K. Zhou et al. / Computer Networks 50 (2006) 37843794
x
q
= 0.002, min
th
= 10, and max
th
= 30. The
default parameters of REM in ns-2 are used, they
are c = 0.001, a = 0.1, / = 1.001, and b = 20.
From Fig. 12, we can see that NLRED has the
highest overall throughput, whereas GRED is the
lowest. It is interesting to see a short concave phase
when the trac is changed from 10 ows to 40
ows. However, as soon as the number of ows is
larger than 40, the throughput for NLRED quickly
converges to the link bandwidth. Besides, during the
concave range, the throughput of NLRED is still
always higher than GRED. Fig. 13 shows the corre-
sponding average queue size of using GRED, REM,
and NLRED. By steering the queue around a target
length, REM suers the low throughput when trac
load is extremely light (less than 5 ows) and extre-
mely high. When N > 60, the throughput of REM is
unstable and drops as N increases.
4.3. Experiment 3
Since Misra et al. [15] indicated that packet size
aects the performance of AQM schemes, in this
experiment (again based in Fig. 5), we test and com-
pare the packet size sensitivity of GRED, REM,
and NLRED. REM is congured to work in byte
mode because packet mode shows extremely poor
180 182 184 186 188 190 192 194 196 198 200
0
10
20
30
40
50
60
70
80
90
100
Time (seconds)
Q
u
e
u
e

l
e
n
g
t
h

(
p
a
c
k
e
t
s
)
EWMA Average queue
Instantaneous queue
Fig. 11. Change in queue occupancy when GRED is used with
N = 100 ows, max
p
= 0.02, x
q
= 0.002, min
th
= 10, max
th
= 30.
180 182 184 186 188 190 192 194 196 198 200
10
20
30
40
50
60
70
80
90
100
Time (seconds)
Q
u
e
u
e

l
e
n
g
t
h

(
p
a
c
k
e
t
s
)
EWMA Average queue
Instantaneous queue
Fig. 10. Change in queue occupancy when NLRED is used with
N = 100 ows, max
p
= 0.02, x
q
= 0.002, min
th
= 10, max
th
= 30.
0 20 40 60 80 100 120
90
91
92
93
94
95
96
97
98
99
100
Flows
T
h
r
o
u
g
h
p
u
t
/
B
a
n
d
w
i
d
t
h

(
%
)
GRED
REM
NLRED
Fig. 12. Throughput vs. ow number: GRED, REM and
NLRED.
Flows
0 20 40 60 80 100 120
5
10
15
20
25
30
35
40
A
v
e
r
a
g
e

q
u
e
u
e

l
e
n
g
t
h

(
p
a
c
k
e
t
s
)
GRED
REM
NLRED
Fig. 13. Average queue size vs. ow number: GRED, REM and
NLRED.
K. Zhou et al. / Computer Networks 50 (2006) 37843794 3791
performance.
1
(We believe the error in arrival rate
estimation is the reason for such poor performance
with packet mode REM.) However, the queue
length of byte mode REM cannot be directly com-
pared with the results of other schemes, because it
uses bytes as unit whereas others use packets. To
solve the problem, we normalize the queue length
of byte mode REM to use packets as unit. The con-
version assumes all the packets are with the same
size as the referenced packet size.
We simulate 50 long-lived FTPs. For each AQM
algorithm, we conduct a set of simulations with the
packet size ranging from 100 bytes to 2500 bytes.
Fig. 14 shows the throughput against packet size.
We can see that NLRED is least sensitive to the
packet size and therefore is better than both GRED
and REM.
5. Conclusions
In this paper, we proposed a new active queue
management scheme called Nonlinear RED
(NLRED). NLRED is the same as the original
RED except that the linear packet dropping proba-
bility function is replaced by a nonlinear quadratic
function. While inheriting the simplicity of RED,
NLRED was shown to outperform RED as well
as REM and some of its variants. In particular,
NLRED is less sensitive to parameter settings, has
a more predictable average queue size, and can
achieve a higher throughput. We credit the above
performance gain to the idea of encouraging the
router to operate over a range of queue sizes accord-
ing to trac load instead of at a xed one. This is
realized in NLRED by using a gentle packet drop-
ping probability at the onset of the congestion,
and a much more aggressive dropping probability
when the congestion becomes more pronounced.
References
[1] L. Zhang, S. Shenker, D.D. Clark, Observations on the
dynamics of a congestion control algorithm: the eects of
two-way trac, in: Proceeding of ACM SIGCOMM 91, the
Conference on Communications Architecture and Protocols,
Zurich, Switzerland, September 0306, 1991, pp. 133147.
[2] R. Mahajan, S. Floyd, D. Wetherall, Controlling high-
bandwidth ows at the congested router, in: Proceeding of
Ninth International Conference on Network Protocols
(ICNP01), Riverside, California, USA, November 1114,
2001, pp. 192201.
[3] S. Floyd, J. Mahdavi, M. Mathis, et al., An extension to the
selective acknowledgement (SACK) option for TCP, in:
RFC 2883, 2000.
[4] K. Xu, N. Ansari, Stability and fairness of rate estimation
based AIAD congestion control in TCP, IEEE Communi-
cations Letters 9 (2005) 378380.
[5] I.S. Institute, Transmission control protocol, in: RFC793,
1981.
[6] S. Floyd, K. Fall, Promoting the use of end-to-end conges-
tion control in the Internet, IEEE/ACM Transactions on
Networking 7 (1999) 458472.
[7] B. Braden, D. Clark, J. Crowcroft, et al., Recommendations
on queue management and congestion avoidance in the
Internet, in: RFC2309, April 1998.
[8] S. Floyd, V. Jacobson, Random early detection gateways for
congestion avoidance, IEEE/ACM Transactions on Net-
working 1 (1993) 397413.
[9] Cisco Systems, Congestion Avoidance Overview. Available
from: <http://www.cisco.com/univercd/cc/td/doc/product/
software/ios120/12cgcr/qos_c/qcpart3/qcconavd.htm>.
[10] C. Brandauer, G. Iannaccone, C. Diot, et al., Comparison of
tail drop and active queue management performance for
bulk-data and Web-like Internet trac, in: Proceeding of
Sixth IEEE Symposium on Computers and Communications
(ISCC01), Hammamet, Tunisia, July 0305, 2001, pp. 122
129.
[11] W.-C. Feng, D.D. Kandlur, D. Saha, et al., A self-cong-
uring RED gateway, in: Proceeding of Eighteenth Annual
Joint Conference of the IEEE Computer and Communica-
tions Societies, (INFOCOM99), vol. 3, New York, NY,
USA, March 2125, 1999, pp. 13201328.
[12] T.J. Ott, T.V. Lakshman, L.H. Wong, SRED: stabilized
RED, in: Proceeding of Eighteenth Annual Joint Conference
of the IEEE Computer and Communications Societies,
0 500 1000 1500 2000 2500
84
86
88
90
92
94
96
98
100
Packet size (bytes)
T
h
r
o
u
g
h
p
u
t
/
B
a
n
d
w
i
d
t
h

(
%
)
GRED
REM (byte mode)
NLRED
Fig. 14. Throughput vs. packet size.
1
Packet mode REM is processed on a packet basis, i.e. all the
variables in (2) use packets as unit. For byte mode REM, the
queue length q(k) and arrival rate k(k) are measured in bytes;
however, the nal result r is required to be in packets (as in (3)),
so a variable called referenced packet size is introduced. All the
variables in bytes are then mapped into packets by simply
dividing by the referenced packet size which is 500 bytes in ns-2
by default.
3792 K. Zhou et al. / Computer Networks 50 (2006) 37843794
(INFOCOM 99), vol. 3, New York, NY, USA, March 21
25, 1999, pp. 13461355.
[13] M. May, J.-C. Bolot, A. Jean-Marie, et al., Simple perfor-
mance models of dierentiated services schemes for the
Internet, in: Proceeding of Eighteenth Annual Joint Confer-
ence of the IEEE Computer and Communications Societies,
(INFOCOM 99), vol. 3, New York, NY, USA, March 125,
1999, pp. 13851394.
[14] B. Zheng, M. Atiquzzaman, DSRED: an active queue
management scheme for next generation networks, in:
Proceeding of 25th Annual IEEE Conference on Local
Computer Networks (LCN00), Tampa, Fl, USA, November
810, 2000, pp. 242251.
[15] V. Misra, W.-B. Gong, D. Towsley, Fluid-based analysis of a
network of AQM routers supporting TCP ows with an
application to RED, ACM SIGCOMM Computer Commu-
nication Review 30 (2000) 151160.
[16] S. Athuraliya, S.H. Low, V.H. Li, et al., REM: active queue
management, IEEE Network 15 (2001) 4853.
[17] S. Floyd, R. Gummadi, S. Shenker, Adaptive RED: An
algorithm for increasing the robustness of REDs active
queue management, 2001. Available from: <http://www.
icir.org/oyd/papers/adaptiveRed.pdf>.
[18] W. Feng, D. Kandlur, D. Saha, et al., Blue: An alternative
approach to active queue management, in: Proceeding of the
11th International Workshop on Network and Operating
Systems Support for Digital Audio and Video (NOSSDAV
01), Port Jeerson, NY, USA, June 2526, 2001, pp. 4150.
[19] C.V. Hollot, V. Misra, D. Towsley, et al., Analysis and
design of controllers for AQM routers supporting TCP
ows, IEEE Transactions on Automatic Control 47 (2002)
945959.
[20] S.S. Kunniyur, R. Srikant, An adaptive virtual queue (AVQ)
algorithm for active queue management, IEEE/ACM Trans-
action on Networking 12 (2004) 286299.
[21] L. Zhu, N. Ansari, Local stability of a new adaptive queue
management (AQM) scheme, IEEE Communications Letters
8 (2004) 406408.
[22] T. Bonald, M. May, J.-C. Bolot, Analytic evaluation of RED
performance, in: Proceeding of the Nineteenth Annual Joint
Conference of the IEEE Computer and Communications
Societies (INFOCOM00), vol. 3, Tel Aviv, Israel, March
2631, 2000, pp. 14151424.
[23] Y. Gao, J.C. Hou, A state feedback control approach to
stabilizing queues for ECN-enabled TCP connections, in:
Proceeding of the Twenty-Second Annual Joint Conference
of the IEEE Computer and Communications Societies
(INFOCOM03), vol. 3, San Francisco, CA, USA, March
30April 3, 2003, pp. 23012311.
[24] S. Bohacek, K. Shah, G.R. Arce, et al., Signal processing
challenges in active queue management, IEEE Signal Pro-
cessing Magazine 21 (2004) 6979.
[25] S. Kunniyur, R. Srikant, Analysis and design of an adaptive
virtual queue (AVQ) algorithm for active queue manage-
ment, ACM SIGCOMM Computer Communication Review
31 (2001) 123134.
[26] S. Floyd, Recommendation on using the gentle_variant of
RED, March 2000. Available from: <http://www.icir.org/
oyd/red/gentle.html>.
[27] UCN/LBL/VINT, Network Simulator NS2. Available
from: <http://www-mash.cs.berkeley.edu/ns>.
[28] D. Lapsley, S. Low, Random early marking for Internet
congestion control, in: Proceeding of Global Telecommuni-
cations Conference (GLOBECOM 99), vol. 3, Rio de
Janeiro, Brazil, December 59, 1999, pp. 17471752.
Zhou Kaiyu received the M.S. and B.S.
degree in electronics from the University
of Electronic Science and Technology of
China, Cheng Du, China, in 1994 and
1997, respectively. He is currently
working toward the Ph.D. degree in the
Department of Electrical and Electronic
Engineering at the University of Hong
Kong, Hong Kong, China.
His research focuses on transport layer
protocol analysis and improvement,
Internet congestion control and avoidance.
Kwan L. Yeung was born in 1969. He
received his B.Eng. and Ph.D. degrees in
Information Engineering from The Chi-
nese University of Hong Kong in 1992
and 1995, respectively. He joined the
Department of Electrical and Electronic
Engineering, The University of Hong
Kong in July 2000, where he is currently
an Associate Professor. Before that, he
has spent ve years in the Department of
Electronic Engineering, City University
of Hong Kong as an Assistant Professor. During the summer of
1993, He served with the Performance Analysis Department,
AT&T Bell Laboratories (now Bell Labs, Lucent Technologies),
Holmdel, USA, as a Member of Technical Sta-I. His research
interests include next-generation Internet, active queue manage-
ment, packet switch/router design, all-optical networks and
wireless data networks. He is a regular technical reviewer for many
international journals and conferences, and a regular reviewer of
CERG grant applications. He has published about 100 papers in
international journals and conferences. He was ranked First Place
in 1993 IEEE (HK) Student Paper Contest (postgraduate). He has
also been listed in Marquis Whos Who in the World and Whos
Who in Science and Engineering. In the academic year of 1998
1999, he was recognized as an Outstanding Teacher in the
Department of Electronic Engineering, City University of Hong
Kong. He becomes a senior member of IEEE since 1999.
Victor O.K. Li received SB, SM, EE and
ScD degrees in Electrical Engineering
and Computer Science from the Mas-
sachusetts Institute of Technology,
Cambridge, Massachusetts, in 1977,
1979, 1980, and 1981, respectively. He
joined the University of Southern Cali-
fornia (USC), Los Angeles, California,
USA in February 1981, and became
Professor of Electrical Engineering and
Director of the USC Communication
Sciences Institute. Since September 1997 he has been with the
University of Hong Kong, Hong Kong, where he is Chair
K. Zhou et al. / Computer Networks 50 (2006) 37843794 3793
Professor of Information Engineering at the Department of
Electrical and Electronic Engineering. He has also served as
Managing Director of Versitech Ltd. (http://www.versi-
tech.com.hk/), the technology transfer and commercial arm of
the University, and on various corporate boards. His research is
in information technology, including all-optical networks, wire-
less networks, and Internet technologies and applications. He is
a Principal Investigator of the Area of Excellence in Information
Technology funded by the Hong Kong government. Sought by
government, industry, and academic organizations, he has lec-
tured and consulted extensively around the world. He chaired
the Computer Communications Technical Committee of the
IEEE Communications Society 19871989, and the Los Angeles
Chapter of the IEEE Information Theory Group 19831985. He
co-founded the International Conference on Computer Com-
munications and Networks (IC3N), and chaired its Steering
Committee 19921997. He also chaired various international
workshops and conferences, including, most recently, IEEE
INFOCOM 2004. He has served as an editor of IEEE Network,
IEEE JSAC Wireless Communications Series, and Telecommu-
nication Systems. He also guest edited special issues of IEEE
JSAC, Computer Networks and ISDN Systems, and KICS/
IEEE Journal of Communications and Networking. He is now
serving as an editor of ACM/Kluwer Wireless Networks and
IEEE Communications Surveys and Tutorials. He has been
appointed to the Hong Kong Information Infrastructure Advi-
sory Committee by the Chief Executive of the Hong Kong
Special Administrative Region. He is a part-time member of the
Central Policy Unit of the Hong Kong Government. He also
serves on the Innovation and Technology Fund (Electronics)
Vetting Committee, the Small Entrepreneur Research Assistance
Programme Committee, the Engineering Panel of the Research
Grants Council, and the Task Force for the Hong Kong Aca-
demic and Research Network (HARNET) Development Fund
of the University Grants Committee. He was a Distinguished
Lecturer at the University of California at San Diego, at the
National Science Council of Taiwan, and at the California
Polytechnic Institute. He has also delivered keynote speeches at
many international conferences. He has received numerous
awards, including, most recently, the Changjiang Chair Profes-
sorship from the Ministry of Education, China, the UK Royal
Academy of Engineering Senior Visiting Fellowship in Com-
munications, the Outstanding Researcher Award of the Uni-
versity of Hong Kong, the Croucher Foundation Senior
Research Fellowship, and the Bronze Bauhinia Star, Govern-
ment of the Hong Kong Special Administrative Region, China.
He was elected an IEEE Fellow in 1992. He is also a Fellow of
the HKIE and the IAE.
3794 K. Zhou et al. / Computer Networks 50 (2006) 37843794