IMPROVEMENT OF VOIP QUALITY BY PACKET DROPPING
IN ADSL ROUTERS
Qin Dai, Matthias Baumann
Technische Universität Dresden, Dresden, Germany
Ralf Lehnert
Technische Universität Dresden, Dresden, Germany
Keywords: VoIP, QoS, ADSL Downlink, Queueing Delay, End-to-End Delay, Loss Bursts.
Abstract: Packet dropping is known as a simple mechanism to control TCP traffic. In this paper, TCP packet dropping
is introduced in the egress router of an ADSL downlink. The aim is to improve the quality of VoIP
connections that compete with TCP applications in downlink direction. The ADSL downlink buffer is
assumed to operate as simple FCFS queue. Different simulations have been conducted that evaluate the
mechanism in two scenarios. Firstly, the long-term impact of the mechanism both on VoIP application and
TCP applications is investigated. Secondly, with more realistic network settings, the effectiveness of the
mechanism for a short-time real speech is evaluated. The speech’s PESQ estimate is used to assess the
service quality. The results indicate that in both cases packet dropping can improve the VoIP quality.
However, the required high dropping ratio can result in TCP traffic bursts and therefore unstable VoIP
quality as well as bad TCP performance.
1 INTRODUCTION
As one kind of real-time applications, VoIP (voice
over IP) has its QoS (quality of service)
requirements on delay, delay variation and packet
loss. Among them, delay related measurements are
most critical. The first big problem caused by delay
is the talker overlap. In a two-way conversation it is
hard to continue when both sides start talking at the
same time. The presence of voice echo also has a
significant impact on delay sensitivity. Moreover
those packets which have been delayed for “too”
long time will be discarded by the de-jittering buffer
in the receiver.
VoIP packet delay, which can be caused by
different reasons, can be characterized into two basic
types: the fixed delay and the variable delay. Fixed
delay includes coding / decoding delay, processing
delay, propagation delay, serialization delay etc. All
of these delays are fixed according to the applied
technologies, e.g. the algorithms for coding /
decoding, the bit rate for the packet serialization,
and the propagation delay itself. Variable delay
mainly is caused by varying queue occupations in
network nodes which largely depend on the situation
of competing traffic streams. The queueing delay of
a VoIP packet especially can become serious when
entering a bottleneck link with fast packet stream
aggregation at the input side and slow release at the
output side.
In recent years, ADSL (Asymmetric Digital
Subscriber Line) technology has become one of the
most popular access solutions for broadband
networks. In ADSL access network, voice and data
traffic will be multiplexed / demultiplexed in its
edge node to the IP/ATM network named DSLAM
(Digital Subscriber Line Access Multiplexer). From
the comparatively low speed in the uplink (typically
128Kbit/s), one can expect that QoS problems
mainly arise in this direction, and most of the
research focuses on the uplink problem, e.g.
(Orozco_barbosa, Siddiqui, Yongacoglu, 2002).
Although the downlink speed of ADSL may be quite
large (up to 8 Mbit/s), the considerable bandwidth
difference between core and access network leads to
queueing of downstream packets in the DSLAM. If
the IP network operator does not support different
service types for VoIP and data applications, FCFS
269
Dai Q., Baumann M. and Lehnert R. (2007).
IMPROVEMENT OF VOIP QUALITY BY PACKET DROPPING IN ADSL ROUTERS.
In Proceedings of the Second International Conference on Signal Processing and Multimedia Applications, pages 265-272
DOI: 10.5220/0002135502650272
Copyright
c
SciTePress
(First Come First Serve) queueing may be necessary.
In this paper, we consider such a scenario: different
traffic streams including VoIP, TCP (Web and FTP
traffic) go through the core IP network and pass the
access node DSLAM. The link between DSLAM
and ADSL modem becomes the bottleneck in the
downlink of the communication. The FCFS
queueing in DSLAM leads to unacceptable delay for
VoIP packets.
ITU-T recommendation G.114 states that the
one-way VoIP communication delay should not
exceed 150 ms. We assume a voice communication
with ADSL access on both end points and consider
the downlink performance in one of the points. We
neglect the delay in the Internet and assume that a
non-adaptive de-jittering buffer with the size of 150
ms is used in the receiver side. For ADSL links with
1 MBit/s downlink and 128 kBit/s uplink bit rate, up
to 100 ms of the total delay budget should be
reserved for the remote uplink part of the VoIP
connection. In order to restrict the downlink delay to
the remaining budget, the number of packets queued
by TCP connections has to be controlled.
As one kind of TCP traffic control mechanisms,
packet dropping can effectively limit the TCP
sending rate and has been implemented with
different intentions. To avoid network congestion
and improve the overall performance, Random Early
Detection (RED) uses a linearly increasing function
of the queue length to probabilistically drop
incoming packets with the expectation that the
remote sender can slow down its transmission rate
(Floyd, Jacobson, 1993). (Shyu, Chen, Luo, 2002)
introduces an adaptive dropping mechanism in the
router for Internet congestion control. In this paper,
we implement packet dropping in the ADSL router
at the user’s side and expect that the number of TCP
packets delaying VoIP packets in the DSLAM queue
can be controlled. We investigate the influence of
the mechanism both on the behaviour of the TCP
control loop and on the VoIP application. In our
simulations, we firstly investigate both VoIP and
TCP quality on packet level by using a long time
VoIP session. In the second simulation, we use a
short-time real speech as the VoIP streaming since
the real conversation consists of short speech
sections and silence periods. Additionally, higher
bandwidths both in up and down links are reserved
for the ADSL user in this simulation. A perceptual
evaluation of the VoIP quality will be conducted by
using the PESQ model.
The remainder of this paper is organized as
follows. In Section 2, the packet dropping
mechanism is analyzed and discussed. Later the
PESQ model is introduced. In Section 3, the
simulation models and the metrics for performance
evaluation in two simulations are introduced. The
numerical results are presented in Section 4. Finally
the conclusions are drawn.
2 BACKGROUND
2.1 TCP Packet Dropping
As already mentioned in the last section, it is
necessary to limit the number of packets a TCP
connection can place in the DSLAM buffer. TCP is
a sliding window-based protocol. The TCP sender
restricts its window size according to the minimum
of the receiver’s buffer size and the congestion
window size. In TCP, packet loss is considered as a
sign of network congestion. Reacting on packet loss,
the sender reduces its congestion window to half of
the old value during “Fast Retransmission and
Recovery” phase to reduce network congestion. In
case of retransmission time-out, the window is even
closed to one packet. Thus it seems feasible to avoid
large window sizes by introducing “artificial” packet
losses in the downstream of active TCP connections.
If packets of a single connection are dropped
periodically, the TCP window size cannot exceed a
certain threshold, thus effectively limiting the
maximum number of packets that are placed in the
DSLAM buffer. One drawback of this approach is
that the dropping ratio should depend on the network
parameters, e.g. the number of TCP connections and
the lowest round-trip time (RTT), which can lead to
very low link utilization and bad perceived
performance e.g. for Web applications. Our
simulation results will further illustrate the problem.
Besides it might appear counterintuitive to drop
packets already transmitted successfully over the
ADSL link. In spite of the above disadvantages, the
approach‘s simplicity could lead to an almost gratis
implementation still offering a sensible QoS
advantage.
2.2 PESQ Model
As one objective method for speech quality
assessment, the approach Perceptual Evaluation of
Speech Quality (PESQ) is standardised by ITU-T as
recommendation P.862. Figure 1 shows the basic
philosophy used in PESQ (ITU-T, 2001). Firstly, a
series of delays between original input and degraded
output are computed by the time alignment module.
Secondly, based on this set of delays PESQ
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
270
compares the original signal with the aligned
degraded one by using a perceptual model. The key
to this process is transformation of both the original
and degraded signals to an internal representation
that is analogous to the psychophysical
representation of audio signals in the human
auditory system. The differences in internal
representation between the original signal and the
degraded one are computed in the cognitive model.
Finally, an objective listening quality MOS (mean
opinion score) is given. PESQ is believed to be able
to predict subjective quality with good correlation in
a very wide range of conditions, that may include
coding distortions, errors, noise, filtering, delay and
variable delay (Rix, Hollier, Hekstra, Beerends,
2002). The average correlation coefficient between
PESQ and the subjective scores is indicated as 0.935
in (ITU-T, 2001) from 22 known ITU benchmark
experiments. (ITU-T, 2001)
Table 1 shows the ACR (Absolute Category
Rating) opinion scale used in the development of
PESQ (Beerends, Hekstra, Rix, Hollier, 2002). The
speech quality is graded from bad to excellent with
the PESQ score increases from 1 to 5. Since a
comparable voice quality is always expected when
the user switches from the traditional telephone to
VoIP, a speech with low PESQ value can not be
acceptable. Different codecs can introduce different
level of degrading to the voice. (Markopoulou,
Tobagi, Karam, 2003) indicates that the ideal MOS
value of a speech coded with G.729 is around 4. The
PESQ value of 3.5 is taken as the criterion for
judging the quality of speech in our second
simulation, with which the quality of the speech is
considered between fair to good as shown in the
table.
Table 1: ACR listening quality opinion scale used in the
development of PESQ.
Quality of the speech Score
Excellent 5
Good 4
Fair 3
Poor 2
Bad 1
Figure 1: Overview of the basic philosophy used in PESQ
(ITU-T, 2001).
VoIP phone
VoIP
downlink
delay
TCP applications
(web and FTP
clients
)
Ethernet
Web servers
TCP Internet
access node
FTP server
ATM
VoIP
DSLAM
ADSL downlink ADSL uplink
DSLAM buffer
(FCFS)
PC
ADSL router
VoIP
ATM
Figure 2: System model (ATM links have a bit rate of 150 Mbit/s).
IMPROVEMENT OF VOIP QUALITY BY PACKET DROPPING IN ADSL ROUTERS
271
3 SIMULATIONS AND
PERFORMANCE METRICS
3.1 Simulation I
Figure 2 shows the system model adopted in this
simulation. Firstly, we investigate the VoIP quality
using only Web traffic as background load. In a Web
application, usually several short-lived TCP
connections run simultaneously. This leads to bursty
traffic, since the slow start at the beginning of a TCP
connection results in steep packet rate increase. In
our simulation, we use two concurrent Web sessions
to intensify the traffic burstiness. Within each
session, documents from a server farm of 100 Web
servers are downloaded in a random fashion. These
servers are connected to the Internet part of the
network model with link delays ranging from 15 to
35 ms. The traffic model for the Web applications is
taken from (Mah, 1997). It specifies distributions for
the distance between Web requests (exponential), for
the number of objects per main page (Pareto), for the
distance between object requests in a page
(exponential), and for the object sizes (Pareto). In
order to assess the influence of the approach, we
measure mean response times for Web pages with a
total size of less than 30 Kbytes (the mean bit rate of
the Web applications is rather small).
Secondly, only a greedy FTP connection is
combined with the VoIP traffic. This allows us to
evaluate the possible download rate for a bulk data
transfer. Different network delays of 1, 10 and
100 ms for the FTP connection shall yield further
insight into the performance of the investigated
approach.
The VoIP connection is modelled as bidirectional
flow with isochronous payload of 20 bytes every
20 ms (corresponding to G.729A). Including all
overheads on the ADSL link (RTP, UDP, IP,
PPPoE, AAL, ATM), this leads to an effective
amount of 159 bytes per voice packet. In the uplink
queue of the ADSL router, scheduling with absolute
priority of VoIP packets is assumed.
For this simulation, only packet-level performance
metrics have been used to yield an estimate of the
eventual speech quality. Recent investigations, e.g.
(Markopoulou, Tobagi, Karam, 2003), (James,
Chen, Garrison, 2004), find that the mean packet
loss ratio should not exceed 1%, and that packet
losses in clusters or bursts can be compensated to
different degrees by PLC (packet loss concealment)
algorithms of the voice decoder. According to this,
we assess the QoS of VoIP using the average packet
loss ratio as well as the distribution of loss burst
lengths. Two losses are considered to be part of the
same burst if less than 10 packets are transmitted
successfully in between. As the robustness of PLC
algorithms may vary, we plot the frequency of loss
bursts longer then 2, 5 and 8 packets in the Figures 4
and 5 below.
In the applied network model, all buffer sizes are
set to very large values. Thus, packet losses for the
VoIP connection only are caused by excessive delay
in the DSLAM queue and subsequent packet
dropping in the de-jittering buffer of the VoIP
phone. In this paper, we only consider a very simple
algorithm for the de-jittering buffer. All packets
arriving “in time” will be buffered till a threshold
time before being played out. Conversely, any
packets that arrive late will be dropped by the buffer.
The numerical results below assume a delay budget
of 50 ms between arrival at the DSLAM buffer and
packet delivery to the decoder.
All simulations have been performed for 100
sub-runs each of 10 minutes model time, the
confidence level is 95%. The simulations are
realized in ns-2 (http://www.isi.edu/nsnam/ns/).
3.2 Simulation II
Figure 3 illustrates the experimental structure. The
source signal holding about 20 seconds speech will
be coded and packtized into VoIP packets in the
VoIP sender. After transversing through the network
introduced in Simualtion I, the speech signal will be
recovered from those packets with the help of the
data extractor and the decoder. Finally, the PMOS
(PESQ MOS) value will be computed in PESQ by
comparing the degraded speech file with the original
one.
The increasing interest on applications like VoD
(Video on Demand) and interactive Internet gaming
leads to rising bandwidth demand. Correspondingly
the ISPs also offer higher bandwidth to their
customers. This higher bandwidth might also solve
the QoS problem of some applications having
comparatively low bandwidth requirements like
VoIP. In this simulation, besides the normal rate as
set in simulation I, doubled bandwidths are set to
both of ADSL up and down links for investigation.
The network model introduced in simulation I is
adopted. Two Web sessions running simultaneously
are set to the background load. The coding rate of
G.729 is set to 8Kbit/s which leads to 20 Bytes
payload for every 20 ms. We turn off the VAD
(Voice Activity Detection) function in codec to
maximize the bandwidth requirement of VoIP
application. For estimating voice quality, the mean
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
272
Figure 3: Structure in simulation II.
packet loss ratio and the number of the loss burst in
20 s, as well as the mean PESQ value and the bad
PESQ ratio are used. The latter measure indicates
the relative frequency of PESQ values below 3.5. In
simulation I, the loss distance (the difference in
sequence numbers between two successively lost
packets) is defined as 10 under the assumption that
PLC is disabled in the decoder. According to (Levy,
Zlatokrilov, 2006), small values of loss distance can
be used when FEC (Forward Error Correction) or
PLC mechanisms are enabled. Based on the
assumption that the codec G.729 only can repair 30
to 40 ms losses (Perkins, Hodson, Hardman, 1998),
the number of loss bursts with length great than 2
(20 ms per packet) is depicted in Figure 6 below. In
both cases we assume that 100 ms delay budget
should be reserved for the uplink of the other side
and use a non-adaptive de-jittering buffer with size
of 150 ms in the receiver. The simulations have been
performed for 10x40 times; within each time the
whole transmission of the test speech file is
completed. In order to clearly separate to sub-runs, a
silence period of 300 ms is introduced between two
speech transmissions.
4 SIMULATION RESULTS AND
COMPARISONS
4.1 Simulation I
Figure 4 shows the VoIP quality in ADSL
downlinks with different delay budgets when
competing with 2 Web sessions. To ensure a VoIP
packet loss ratio below the tolerable value of 1%,
about 500 ms delay budget should be reserved for
the VoIP downlink (between DSLAM and the
receiver) which is far above the limitation.
Figure 5 shows the performance of VoIP and
TCP with packet dropping in ADSL router. As
discussed in Section 2, the introduced dropping ratio
should increase with the number of TCP
connections. In order to reduce the queueing delay
of VoIP packet in DSLAM to an acceptable level of
e.g. 50 ms, the queue length at arrival of a VoIP
packet should not exceed 4 packets (TCP packet
size: 1500 bytes, downlink bit rate: 1 Mbit/s). In the
case of only one connection, the TCP window size
therefore should be limited to 4 packets. With Web
sessions in the network, TCP packets from different
connections in a session can arrive simultaneously at
the DSLAM queue and provoke a larger queue
occupation. To ensure that the accumulated queue
length will not exceed 4, the window size for each
TCP connection should be further limited. Hence, a
very high drop ratio should be introduced which
may lead to low link utilization. High packet loss
ratios cause poor TCP behaviour during the start of
the connection and thus result in bad perceived
performance especially for Web applications.
Additionally, the dropping ratio should be adjusted
to a value that ensures the VoIP quality for the
lowest possible packet round-trip time (RTT) of all
TCP connections. In case of actually larger RTTs, a
certain fraction of the TCP packets will be in transit
on links or waiting in queues other than the
DSLAM. Thus, the TCP performance will be worse
than necessary.
As shown in Figure 5 (left side), discarding more
than 5% TCP packets can push the VoIP packet loss
ratio (fraction of late packets) below 1%, while the
frequency of packet loss bursts still can be serious.
We consider that the number of loss bursts with
length greater than 5 packets should not exceed 5
source
G.729
Packet
generator
Data
extractor
G.729
PESQ
degraded
signal
PMOS value
Network model
ns2 Model
IMPROVEMENT OF VOIP QUALITY BY PACKET DROPPING IN ADSL ROUTERS
273
within 10 minutes. To satisfy this requirement, it
follows from the figure that more than 10%
dropping ratio should be introduced. The right side
of Figure 5 demonstrates that then both FTP
throughput and mean Web response time deteriorate
considerably.
4.2 Simulation II
Figure 6 shows the VoIP performance in ADSL
downlinks when competing with 2 Web sessions.
In the case of 1 Mbit/s downlink bit rate, one
may find that our algorithm performances much
better on a short-time speech than on a long speech
(simulation I). The fraction of late VoIP packet can
be controlled below our criteria of 1% with a drop
period of 40 packets, while simulation I led to a
dropping ratio of more than 1/25. As shown in
Figure 6, the mean PESQ value can be quite good
with only 1/90 drop ratio. At the same time, the
mean number of loss bursts in 20 seconds is only
about 1. Nevertheless, more than 10% drop ratio is
required to hold the relative frequency of bad PESQ
values below 1%. This means that the user will not
more than once experience a bad VoIP quality
during 30 minutes.
Table 2 depicts the VoIP parameters with 2
MBit/s downlink bit rate on the ADSL link. Packet
dropping is not applied. The table shows that the
packet-level performance metrics do not indicate
QoS problems. The fraction of late VoIP packets in
20 s is only 1% which is almost within our criterion.
The mean number of loss bursts with burst length
greater than two is 0.4 times in 20 s. But the serious
situation of VoIP application under this scenario can
still be seen from the bad PMOS fraction which is
indicated as 15%. This means that the user will be
dissatisfied with the conversation quality more than
15 times within 33 minutes. The green curve in
Figure 7 addresses this problem more clearly.
According to the bursty background traffic, the VoIP
quality can be very different over time. In the total
100 simulation runs, there are 15 PESQ value drops
below 3.5, two estimates even fall below 3.0. The
worst PESQ value is indicated as 2.75 in Table 3.
Figure 6 indicates that dropping every 30
th
packet could be a good operation point in this
scenario. With this setting, only 0.06% of the VoIP
packets are lost, and the occurrence of bad PESQ
values can be almost suppressed (indicated to be
under 0.8%). However, the QoS can not be
guaranteed, since dropping of TCP packets can
trigger loss bursts of VoIP. When the lost TCP
segment is successfully retransmitted, the receiver
will send back a packet that acknowledges the lost
together with subsequent correctly transmitted
packets. This can lead to a sudden shift of the
transmission window and therefore cause a burst of
packets. This TCP burst can delay a sequence of
VoIP packets, leading to a VoIP loss burst and thus
degrading the voice quality. The sample curve of
PESQ values in Figure 7 highlights this aspect.
Although the PESQ estimate can be improved for
most of the speech intervals, the approach
sometimes can even make worst the speech quality.
This problem seems to be solved by further
increasing the TCP packet dropping ratio and
therefore limiting the TCP window size furtherly.
With a dropping ratio extended to e.g. 1/15,
sustained good PESQ values can be obtained,
varying from 3.71 to 3.75. As seen from simulation
I, this, however, will lead to severely reduced TCP
performance.
Figure 4: Performance of VoIP in ADSL downlinks in simulation I & Figure 7: PESQ values with different algorithms.
1e-06
1e-05
1e-04
0.001
0.01
0.1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
3
6
9
12
15
18
21
24
27
30
Fraction of late VoIP packets
Number of loss bursts in 10 minutes
Delay budget / s
Late VoIP packets
Burst length > 2
Burst length > 5
Burst length > 8
2.5
3
3.5
4
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
PESQ MOS value
Number of iteration
without algorithm
Packet drop (1/30)
Packet drop (1/15)
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
274
Figure 5: Performance of VoIP and TCP with packet dropping in simulation I.
Figure 6: Performance of VoIP with packet dropping in simulation II (single bandwidth).
Table 2: VoIP parameters with double bandwidth without packet dropping.
Fraction of VoIP packets Number of loss bursts (Burst length > 2) Bad PMOS (PMOS < 3.5)
1.01% 0.407 15.8%
Table 3: PESQ values from different algorithms (double bandwidth).
Mean Best Worst Bad fraction (<3.5)
No algorithm 3.654 3.754 2.750 0.15
Packet drop (1/30) 3.740 3.754 3.385 0.01
Packet drop (1/15) 3.754 3.754 3.711 0
5 CONCLUSION
In this paper we investigate the simple approach of
TCP packet dropping to reduce the queueing delay
of VoIP packets in the ADSL downlink under
different TCP traffic conditions. We assess the
performance of the algorithm both in a critical and a
more practical scenario. In the former case VoIP has
to compete for the limited bandwidth with TCP
traffics during a long time. In the latter one, a short-
time speech is investigated with single and double
bandwidth in the ADSL links. Our simulation results
indicate that VoIP suffers from QoS problems even
with high ADSL downlink bandwidth. In both cases,
the VoIP quality can be improved by TCP packet
dropping. The required high dropping ratio can,
1e-06
1e-05
1e-04
0.001
0.01
0.1
10 20 30 40 50 60 70 80 90
0
0.3
0.6
0.9
1.2
1.5
1.8
2.1
2.4
2.7
3
Fraction of late VoIP packets
Number of lossbursts in 20s
Drop period / packet
1Mbit/s: mean loss ratio
2Mbit/s: mean loss ratio
1Mbit/s: number of bursts (burst length > 2)
2Mbit/s: number of bursts (burst length > 2)
0.001
0.01
0.1
1
10 20 30 40 50 60 70 80 90
Bad PMOS fraction
PESQ MOS
Drop period / packet
1Mbit/s: mean PMOS
2Mbit/s: mean PMOS
1Mbit/s: mean bad PMOS fraction (PMOS < 3.5)
2Mbit/s: mean bad PMOS fraction (PMOS < 3.5)
0.001
0.01
0.1
10 15 20 25 30 35 40 45 50
0
5
10
15
20
25
30
35
40
45
50
Fraction of late VoIP packets
Number of loss bursts in 10 minutes
Drop period (in packets)
Fraction of later VoIP packets
Burst length > 2
Burst length > 5
Burst length > 8
0
0.2
0.4
0.6
0.8
1
10 15 20 25 30 35 40 45 50
0
1
2
3
4
5
6
7
8
9
10
FTP throughput / (MBit/s)
Web response time / s
Drop period (in packets)
FTP throughput 1
FTP throughput 10
FTP throughput 100
Web response time
IMPROVEMENT OF VOIP QUALITY BY PACKET DROPPING IN ADSL ROUTERS
275
however, lead to poor TCP behaviour, e.g. low FTP
throughput which yields low link utilization, and
long Web response time. Furthermore, the TCP
packet bursts indirectly caused by the dropping
algorithm can result in unstable VoIP quality.
ACKNOWLEDGEMENTS
This work was supported by the Sphairon Access
Systems GmbH.
REFERENCES
Beerends J. G., Hekstra A. P., Rix A. W., Hollier M. P.,
2002. Perceptual Evaluation of Speech Quality
(PESQ), the new ITU standard for end-to-end speech
quality assessment. Part II – Psychoacoustic model.
Journal of the Audio Engineering Society (J. Audio
Eng. Soc.), 2002.
Floyd S., Jacobson V., 1993. Random Early Detection
gateways for congestion Avoidance. IEEE/ACM
Transaction on Networking, Vol. 1 No. 4, Aug. 1993,
pp.397-413.
ITU-T, 2001. ITU-T Recommendation P.862, 2001.
James J. H.., Chen B., Garrison L., 2004. Implementing o f
VoIP: A Voice Transmission Performance Progress
Report, IEEE Communication Magazine, pp. 36-41,
July 2004.
Levy H., Zlatokrilov H., 2006. The effect of packet
dispersion on voice applications in IP networks.
IEEE/ACM transaction on networking, Vol. 14, No. 2,
April 2006.
Mah B. A., 1997. An Empirical Model of HTTP Network
Traffict, Proc. of IEEE Infocom 1997, April 1997.
Markopoulou A.P., Tobagi F.A., Karam M.J., 2003.
Assessing the Quality of Voice Communications Over
Internet Backbones. IEEE/ACM Transactions on
Networking, vol. 11, no. 5, October 2003, pp. 747-760.
Orozco-Barbosa L., Siddiqui A., Yongacoglu A., 2002.
Efficient integration of voice and data for transmission
over Digtal Subscriber Line (DSL), Proc. of the 2002
IEEE Canadian Coference on Electrical & Computer
Engineering, 2002.
Perkins C., Hodson O., Hardman V., 1998. A survey of
packet loss recovery techniques for streaming audio.
IEEE Network, September/October 1998.
Rix A. W., Hollier M. P., Hekstra A. P., Beerends J. G.,
2002. Perceptual Evaluation of Speech Quality
(PESQ), the new ITU standard for end-to-end speech
quality assessment. Part I – Time alignmentl. Journal
of the Audio Engineering Society (J. Audio Eng.
Soc.), 2002.
Shyu Mei-Ling, Chen Shu-Ching, Luo Hongli, 2003. Per-
class queue management and adaptive packet drop
mechanism for multimedia networking. Proceedings of
the 2003 International Conference on Multimedia and
Expo - Volume 3, 6-9 July 2003.
Spring N. T., Chesire M., Berryman M., Sahasranaman V.,
Anderson T., Breshad B., 2000. Receiver Based
Management of Low Bandwidth Access Links. IEEE
INFOCOM’00, Tel Aviv, Israel, Apr. 2000.
The Network Simulator- ns-2, from
http://www.isi.edu/nsnam/ns/.
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
276