Network condition detection for video transport over wireless Internet

Abstract

Real-time video transport over wireless Internet faces many challenges due to the heterogeneous network environment. A robust network condition classification algorithm using multiple end-to-end metrics and support vector machine (SVM) is proposed to classify different network events and model the transition pattern of network conditions. End-to-end quality-of-service (QoS) mechanisms like congestion control, error control, and power control can benefit from the network condition information and react to different network situations appropriately. The proposed network condition classification algorithm uses SVM as a classifier to cluster different end-to-end metrics such as end-to-end delay, delay jitter, throughput and packet loss-rate for the HDP traffic with TCP-friendly rate control (TFRC), which is used for video transport. The algorithm is also flexible to classify different numbers of states representing different levels of network events such as wireline congestion and wireless channel loss. Simulation results using ns2 show the effectiveness of the proposed scheme

Full text

Page 1

Abstract— Real-time video transport over wireless
Internet faces many challenges due to the heterogeneous
network environment.
A
classification algorithm using multiple end-to-end metrics and
Support Vector Machine (SVM) is proposed to classify different
network events and model the transition pattern of network
conditions. End-to-end Quality-of-Service (QoS) mechanisms like
congestion control, error control, and power control can benefit
from the network condition information and react to different
network situations appropriately. The proposed network
condition classification algorithm uses SVM as a classifier to
cluster different end-to-end metrics such as end-to-end delay,
delay jitter, throughput and packet loss-rate for the UDP traffic
with TCP-friendly Rate Control (TFRC), which is used for video
transport. The algorithm is also flexible to classify different
numbers of states representing different levels of network events
such as wireline congestion and wireless channel loss. Simulation
results using ns2 show the effectiveness of the proposed scheme.
robust network condition
I. INTRODUCTION
Real-time video transport over wireless Internet faces many
challenges due to the heterogeneous environment including
wireline and wireless networks. Figure 1 shows a typical
end-to-end video transport involving wireline and wireless
networks. The video transport may suffer from many problems
such as wireline network congestion and wireless multi-path
fading, which result in high packet loss-rate, and cause severe
video quality degradation. To maintain the optimal video
quality subject to all the
Quality-of-Service (QoS) provisioning for the wireless Internet
is needed and has been an active research area. A general
framework for the end-to-end video transport over the wireless
Internet has been discussed by Zhang et al. [1]. However, to be
effective, the techniques to combat the network impairments
should be dependent on the network conditions. Since the
network condition is dynamic, when there is a change in the
network condition, the end systems should be able to employ
adaptive QoS control mechanisms including congestion
control, error control, and power control to maximize the video
quality. Thus, it is very desirable to be able to identify the
network condition.
Many efforts [2-4] have investigated the use of end-to-end

This work was supported by the Croucher Foundation fellowship from Hong
Kong.
constraints, end-to-end

Figure 1. End-to-end Video Transport involving Wireline and
Wireless Networks

statistics obtained at the receiver to differentiate the loss nature
between the wireline and the wireless channel. Those works
mainly focus on how to design an accurate loss differentiation
algorithm based on a single-metric end-to-end measurement
such as packet inter-arrival time, relative one trip time, or
number of packet loss. Fu et al. [5] used a cascade of decisions
with a single metric in each step to determine the network
condition for supporting TCP-friendly congestion control in
mobile ad-hoc networks. Hidden Markov Model (HMM) is also
applied in differentiating the loss nature between the wireline
congestion and the wireless fading for TCP traffic [6]. However,
the end-to-end loss differentiation based on single metric [2-6]
targets to provide only simple classification between
congestion and wireless channel loss. In practical network
situations, multiple network events can happen simultaneously.
Using single-metric end-to-end measurement may not easily
identify different simultaneous network events. Therefore,
there is a need to design a more sophisticated and general
network condition classification mechanism, which can
provide reliable information for adaptive end-to-end QoS
mechanisms. Besides, it is desirable that the classification
algorithm can also identify different levels of network events
(e.g. the congestion level of the wireline network, and the
degree of the wireless channel loss) and can classify multiple
simultaneous events.
The focus of this paper is to design a robust network condition
classification algorithm using SVM based on multiple metrics,
in order to classify different network events as shown in Figure
2. We focus on the TFRC [7] protocol, which is designed for
supporting video transport over the Internet. We formulate the
network condition classification algorithm using SVM based
on multiple end-to-end metrics and discuss the training
procedure. Using the SVM, it is also flexible to define different
Siu-Ping Chan and Ming-Ting Sun
Department of Electrical Engineering
University of Washington, Seattle 98195, WA, USA
{spchan,sun}@ee.washington.edu
Network Condition Detection for Video
Transport over Wireless Internet
3093
ISCAS 2006 0-7803-9390-2/06/$20.00 ©2006 IEEE

Page 2

numbers of states for different levels of network events like
such as wireline congestion and wireless loss. We also carry out
simulations using ns2 [8] to test the performance of the
proposed algorithm.


Figure 2. Network Condition Classification for Video
Transport
II. NETWORK EVENTS AND END-TO-END METRICS
In this section, we discuss different network events to
classify. We then discuss different end-to-end metrics used in
the proposed network condition classification algorithm with
the SVM, and their importance in different network conditions.
A. Network Events
1) Congestion: Congestion occurs when the buffers in the
routers overflow which results in packet dropping according
to the corresponding buffer management schemes such as
DropTail [9]. Once the packets are dropped, transport
protocols like TCP and TFRC reduce their transmission rate
in order to resolve the congestion situation. Depends on the
traffic load on a router, congestion can be classified into
different levels. For example, we can define four levels of
congestion as non-congestion (NC), slight congestion (SC),
congestion (C), and heavy congestion (HC). Different levels
of congestion have different impacts on the observations of
different end-to-end metrics. For ease of explanation, in our
simulations, the level of congestion is represented by the
number of shared traffic over the same bottleneck link (i.e.
the number of TCP and UDP shared traffic in the
simulations).
2) Wireless Channel Loss: Due to the wireless channel
impairments, packets are corrupted by different channel
fading effects such as multi-path interference and
shadowing. Robust and reliable wireless channel
information is crucial for those end-to-end QoS
mechanisms. In the simulations in this paper, different
levels of wireless channel loss can be defined according to
different wireless channel loss-rates (CLRs) (which
represent the average packet loss-rates in the periods) in the
simulations. For example, we can define three levels of
wireless channel loss as wireless normal (WN), wireless
loss (WL), and severe wireless loss (SWL). Similarly,
different levels of wireless channel loss have different
impacts on the observations of different end-to-end metrics.
The network events of congestion and wireless channel loss
can occur at the same time.
B. End-to-end Metrics and Observations
1) End-to-end Delay. End-to-end delay
the difference between the time
i D is measured as
irt of the packet i being
received by the TFRC receiver and the time
ist
=
of the
−
packet being sent out by the TFRC sender, i.e.
End-to-end delay is increased when congestion happens due
to the additional queueing delay at the bottleneck routers.
2) Delay Jitter: Delay jitter
difference between the end-to-end delays of two
consecutive packets i and j ,
Dtt
=−
and
j
jrs
Dtt
end-to-end delay, delay jitter is increased when congestion
happens, but it is not the case in wireless channel loss. Both
end-to-end delay and delay jitter are effective indicators to
detect the occurrence of congestion events in our algorithm.
3) Throughput: Throughput, Tp , is measured as the
number of bits received per second at the TFRC receiver for
the TFRC traffic. During congestion or wireless channel
loss, the throughput is reduced as the transport protocols,
such as TCP and TFRC, reduce their transmission rate in
order to resolve the congestion situation. It will be shown
that the throughput is also a useful indicator of the
occurrence of those network events. However, only based
on the observation of throughput, it may not be easy to
discriminate the congestion and the wireless loss events.
4) Packet Loss-rate: Packet loss-rate,Pl , is measured as
the rate of packet loss at the TFRC receiver for the TFRC
traffic. Packet loss-rate is an effective metric to detect the
occurrence of wireless loss events in our algorithm.
ii
irs
Dtt
.
iJ is measured as the absolute
||
iij
JDD
=−
, where
ii
irs
j
=−
, respectively. Similar to
III. NETWORK CONDITION DETECTION
SVM have been developed by Boser et al. [10] and Vapnik
[11] to improve accuracy of classifiers in machine learning and
pattern recognition. The advantage of SVM is that it can
achieve high accuracy with relatively small training sets. The
details of SVM have been discussed in [10-11]. To formulate
the network condition classification using the SVM based on
multiple end-to-end metrics, we firstly define the number of
network statesN we want to classify, which depends on the
different end-to-end QoS control strategies to be used. For the
purpose of network condition classification, we can assign
different network states (e.g. non-congestion (NC), slight
congestion (SC), congestion (C) and heavy congestion (HC)) to
different classes in the SVM. We denote the individual network
states as
12
{ ,,,}
N
SS SS
=

.
In order to apply the SVM to estimate the state sequence, we
need to train the SVM. First of all, feature vectors are extracted
from the observation sequences of four end-to-end metrics as
follows:
[ , ,
iii
FeatureVector D J Tp Pl
=
,]1,
ii
i L
≤ ≤
(1)
3094

Page 3

where L is the length of the observation sequence of four
end-to-end metrics, and D , J ,Tp , and Pl are the column
vectors with length L for the end-to-end delay, the delay jitter,
the throughput and the packet loss-rate, respectively. SVM is
then trained by using the training set of the feature vectors
[10-11]. At this step, each feature vector is clustered into N
different classes. After the training, we can obtain the model of
SVM with the number of support vectors and different
coefficients of support vectors, which are used to compute the
optimal hyperplane in order to separate the training samples.
The trained SVM can be used to find the estimated network
state sequence
12
{ ,,,
S S SS
=

vectors extracted from the given testing observation sequences
of different end-to-end metrics. The trained SVM model can be
applied to classify different feature vectors in the testing set
into different estimated network states. Moreover, we apply the
three-point median filter to the estimated network state
sequence,
12
{ ,,,}
N
S S SS
=

, in order to remove the outliners
of the estimated state sequence. We also apply the rule that the
network state transition can only be possible in one single step
only. The motivation is that in practical network situations, it
may not be likely for a non-congestion state to jump into a
heavy congestion state directly. The transition is more likely to
be from non-congestion to slight congestion, and then
congestion and finally heavy congestion. By applying the rule
in state transition, we may be able to model relational and
temporal evolutional pattern of the network state transition.
ˆˆ ˆˆ
}
N
, associated with the feature
ˆ ˆ ˆˆ

Figure 3. Simulation Topology by ns2
IV. RESULTS AND DISCUSSIONS
Simulations by ns2 [8] are performed to show the
performance of our proposed algorithm in different network
conditions. Figure 3 shows the simulation topology and
settings. In these simulations, we used TFRC (with a rate of 0.2
Mbps and a packet-size of 500 bytes) as the video transport
protocol over the wireless Internet. Each link has capacity of
5Mbps, with 2ms of propagation delay. Three shared TCP
traffic (each with a rate of 0.2 Mbps and a packet-size of 500
bytes) and one UDP traffic (with a rate of 0.2Mbps and a
packet-size of 500 bytes) share the same bottleneck link (with a
link capacity of 0.5 Mbps, and 20ms of propagation delay.) to
incur different congestion situations. BS is the BaseStation.
DropTail [9] is used at the bottleneck link routers to manage the
FIFO packet queue. IEEE 802.11 wireless channel with a link
capacity of 11 Mbps is used. The simulation time is 400
seconds. Different wireless channel loss-rates (CLRs) (i.e.
0.02, 0.05, and 0.08) are generated by ns2 for different wireless
channel conditions. The training and the testing periods are
both 400 seconds. In the following simulations, we will first
examine the performance of the proposed algorithm when there
is only one network event occurred in one time period. Multiple
network events occur simultaneously will be investigated in the
next experiment.

A. Detecting Congestion and Wireless Channel loss
In the first set of simulations, we investigate the
performance of the proposed algorithm in classifying two
individual network events (i.e., the congestion and the
wireless channel loss). The events are generated by ns2 and
scheduled to occur in different time periods. In this
simulation, we define those network events and their
corresponding settings as three different states { 1, 2, 3
Table I illustrates the details of different states and their
corresponding simulation settings.

Table I. Different Network States and Settings
State Network State
S1 Non-congestion and
Wireless Normal
S2 Wireless Channel Loss
S SS }.
Simulation Settings
TFRC, and UDP, CLR=0.02
(0s-50s and 350s-400s)
TFRC, and UDP, CLR=0.05
(50s-100s and 300s-350s)
TFRC, UDP, and 3 TCP,
CLR=0.02 (100s-300s)
S3 Congestion

0 50100150200250300 350400
0
0.5
Delay of the received packets (s)
(b) Time /s
050100150 200250300350 400
0
0.02
0.04
Delay Jitter of the received packets (s)
(c) Time /s
050 100 150200 250 300350400
0
0.5
Throughput at the TFRC receiver (Mbps)
(d) Time /s
0 50100150200250 300350400
0
0.5
End-to-end Packet Loss Rate at the TFRC receiver
(e) Time /s
050100150200250 300350 400
0
2
4
Network States (Ground Truth)
(a) Time (s)
0 50 100150200 250 300350400
0
2
4
Estimated Network States
(f) Time /s

Figure 4. Testing results using the proposed algorithm to
classify two individual network events, (f) Precision= 0.9722.

Figure 4 shows the results obtained using the proposed
network condition classification and the observations of the
four end-to-end metrics. It can be observed that both the
end-to-end delay and the delay jitter are effective in
discriminating the wireline network congestion from the
wireless channel loss because there are relatively larger
3095

Page 4

values in those metrics during the congestion than those
during the wireless channel loss. By using the proposed
network condition classification algorithm, we can estimate
the network states as shown in Figure 4 (f) and a precision of
0.9722 can be achieved. The precision is calculated by
counting the number of correctly estimated states with the
ground truth as shown in Figure 4 (a).

Table II. Different Network States and Settings
State Network State
S1 Non-congestion and
Wireless Normal
S2 Slight Congestion and
Wireless Normal
Simulation Settings
TFRC, and UDP, CLR=0.02
(0s-50s and 350s-400s)
TFRC, UDP, and TCP,
CLR=0.02
(50s-100s and 300s-350s)
TFRC, UDP, and 2 TCP,
CLR=0.05
(100s-150s and 250s-300s)
TFRC, UDP, and 3 TCP,
CLR=0.08
(150s-250s)
S3 Congestion and Wireless
Loss
S4 Heavy Congestion and
Severe Wireless Loss
050 100 150200 250300 350400
0
0.2
0.4
Delay of the received packets (s)
(b) Time /s
050 100 150200 250300350400
0
0.02
0.04
Delay Jitter of the received packets (s)
(c) Time /s
0 50100 150200250300 350 400
0
0.5
Throughput at the TFRC receiver (Mbps)
(d) Time /s
0 50 100150 200 250 300 350 400
0
0.5
End-to-end Packet Loss Rate at the TFRC receiver
(e) Time /s
050100150200250300350400
0
2
4
Network States (Ground Truth)
(a) Time (s)
050100150200 250300350 400
0
2
4
Estimated Network States
(f) Time /s

Figure 5. Testing results using the proposed algorithm to
classify multiple simultaneous network events with different
levels, (f) Precision= 0.8246.
B. Detecting Simultaneous Network Events
In this section, we investigate the performance of
classifying multiple simultaneous network events (with
different congestion levels and wireless channel losses
simultaneously). The motivation is that in actual network
situations, multiple network events can happen at the same
time. In this simulation, network events generated by ns2 are
scheduled to occur at the same time period. For the ease of
performing simulations, we consider two different network
events (i.e. congestion and wireless channel loss), each with
different levels (i.e. {Non-congestion, Slight Congestion,
Congestion, Heavy Congestion} and {Wireless Normal,
Wireless Loss, Severe Wireless Loss}). We define four
different states { 1, 2, 3, 4S SS S
}. Table II illustrates the
details of different states and their corresponding simulation
settings. As discussed, our proposed algorithm can be applied
to classify more complex network situations with more levels
of events by defining more states. Figure 5 shows the
observations of different metrics and the estimated network
states using the proposed algorithm. The proposed algorithm
can achieve a precision of 0.8246 as shown in Figure 5 (f).
V. CONCLUSIONS
A robust network condition classification algorithm using the
SVM based on multiple end-to-end metrics is proposed to
classify different network events. The formulation and the
discussions on the training are presented. The proposed
classification algorithm can classify the most likely network
state based on the feature vectors extracted from different
end-to-end metrics observations. Simulations by ns2 have been
performed to investigate the performance of the proposed
algorithm in different network situations such as individual
network events with congestion and wireless channel loss, and
multiple simultaneous network events with different levels.
ACKNOWLEDGEMENT
The authors would like to thanks Professor Ren-Hung Hwang
for his valuable discussions and comments on this paper.
REFERENCES
[1] Q. Zhang, W. Zhu, and Y.Q. Zhang, “End-to-end QoS for Video Delivery
Over Wireless Internet,” Proceedings of the IEEE, vol. 93, no.1, Jan.
2005, pp.123-134.
[2] S. Biaz, and N. Vaidya, “Discriminating congestion losses from wireless
losses using interarrival times at the receiver,” in Proc. of IEEE
Symposium Application-Specific Systems and Software Engineering and
Technology, Richardson, TX, Mar. 1999, pp.10-17.
[3] Y. Tobe, Y. Tamura, A. Molano, S. Ghosh, and H. Tokuda, “Achieving
moderate fairness for UDP flows by path-status classification,” in Proc.
of IEEE LCN, Tampa, FL, Nov. 2000, pp. 252-261.
[4] S. Cen, P.C. Cosman, and G.M. Voelkar, “End-to-end Differentiation of
Congestion and Wireless Loss,” IEEE/ACM Transactions on Networking,
vol.11, no. 5, Oct. 2003, pp. 703–717.
[5] Z. Fu, X. Meng, and S. Lu, “A Transport Protocol for Supporting
Multimedia Streaming in Mobile Ad Hoc Networks,” IEEE Journal on
Selected Areas in Communications, vol. 21, no. 10, Dec. 2003, pp.
1615-1626.
[6] J. Liu, I. Matta, and M. Crovella, “End-to-end Inference of Loss Nature in
Hybrid Wired/Wireless Environment,” in Proc. of WiOpt'03: Modeling
and Optimization in Mobile, Ad Hoc and Wireless Networks, INRIA
Sophia-Antipolis, France, March 2003.
[7] S. Floyd, M. Handley, J. Padhye, and J. Widmer, “Equation-Based
Congestion Control for Unicast Applications,” in Proc. of ACM
SIGCOMM, Stockholm, Sweden, Aug 2000, pp. 43-56.
[8] Network simulator 2 (ns2), http://www.isi.edu/nsnam/ns.
[9] J.F. Kurose, and K.W. Ross, “Computer Networking, A Top-Down
Approach Featuring the Internet, 2nd Edition”, Addison Wesley, 2003.
[10] B.E. Boser, I.M. Guyon, and V. Vapnik, “A Training Algorithm for
Optimal Margin Classifiers,” Proceedings of the 5th Annual ACM
Workshop on Computational Learning Theory, ACM Press, 1992.
[11] V. Vapnik, “Statistical Learning Theory,” Wiley, New York, 1998.

3096