PERFORMANCE OF AUDIO/VIDEO SERVICES ON

CONSTRAINED VARIABLE USER ACCESS LINES

M. Vilas, X. G. Pañeda, D. Melendi, R. Garcia and V. Garcia

Computer Science Department, University of Oviedo, Campues Viesques sn, Xixón/Gijón, Spain

Keywords: User access line, audio/video services, constrained, variable, heavy loaded.

Abstract: Nowadays, it is more and more common for the same access line to be shared among different services and

even among different users. This change in home users’ behaviour, that has given rise to resource

consumption close to the maximum available in user access lines, is mainly due to the increase in subscriber

access capabilities which have taken place in the last few years. At the same time, contracts fulfilled by

customers and network operators only provide guarantees for a reduced percentage of the maximum

download/upload capacity of the line. In this paper, a study of the effects on streaming services caused by

variations on the access line and by the traffic of other services is carried out. One of the main conclusions

of the paper is that the delivery rate of UDP streaming sessions is mainly guided by the quality of the

contents and does not consider the congestion in the network. For this reason, a method for delivery rate

estimation for UDP streaming sessions is presented.

1 INTRODUCTION

The technological evolution of user access lines

which has occurred in the last ten years has led to

deep changes in user behavior. The number of

broadband access lines has increased and

audio/video streaming services, digital newspapers,

virtual offices, email, chats, p2p, etc are services

which are progressively more and more common.

In the contracts fulfilled by customers and

network operators, only a reduced percentage of the

maximum download/upload capacity of the line is

guaranteed, sometimes around 10% of the

maximum. Thanks to this margin, network operators

can reach a balance between investment in resources

on access networks and the level of service provided

to end users. Depending on the number of users

active in each access segment the capacity of these

access lines can vary.

Another parameter that can affect user experience

is the congestion of the access line. It is common to

find users that are sharing files using p2p, while

downloading PDF files using http and meanwhile

accessing an audio/video service looking for the

cinema trailers of next week’s premieres.

For these reasons, it is very interesting to analyze

the effects of bandwidth restrictions on streaming

technology, the effects of other types of user traffic

over streaming services and vice versa. Using this

information, content managers and network

operators can clearly establish the minimum service

levels. In this way, questions such as how much

quality can be offered to a streaming session in the

most loaded hours of the day and how the quality of

streaming sessions is affected can be answered.

In this paper an analysis of the effects on

streaming services of changes in the delivery rate of

access lines of cable network users is performed.

This study is carried out considering another traffic

sharing the bandwidth of the line causing resource

utilization in the user access line close to 100%.

The rest of the paper is organized as follows. In

section 2, previous work in the same field is

analyzed. After that, the test-bed for performance

evaluation is described. In Section 4 the effect of

user access line variations with constrained

bandwidth over streaming services is described.

Based on the results of previous sections, an

estimator of available bit rate that improves the

performance on variable congested user access lines

is presented in section 5. Finally, Conclusions and

Future Work are presented.

Vilas M., G. Pañeda X., Melendi D., Garcia R. and Garcia V. (2007).

PERFORMANCE OF AUDIO/VIDEO SERVICES ON CONSTRAINED VARIABLE USER ACCESS LINES.

In Proceedings of the Second International Conference on Signal Processing and Multimedia Applications, pages 22-27

DOI: 10.5220/0002134900220027

 SciTePress

2 PREVIOUS WORK

The analysis of the traffic generated by some of the

most common streaming platforms, Real Networks

and Windows Media, has been treated in Li,

Claypool & Kinicki (2002) and Kuang & Carey

(2002). These papers analyzed packet inter arrival

times, packet sizes, and the generated packet rate at

different levels.

The co-existence of streaming traffic and other

types of applications has been intensively researched

previously. In Chung & Claypool (2006), the authors

deploy a study of the fairness of Real Networks

streaming flows delivery rate consumption when

they have to share resources with other TCP flows.

The main conclusion of this study is that Real

Networks only present a TCP friendly behavior

when the encoding quality was less than the fair

share of the capacity. In Boyden, Mahanti &

Williamsom (2005), the authors highlight that the

non-TCP friendly nature of Real Networks streaming

was increased when contents are delivered using

Turboplay (RealNetworks, n.d.). In Doshi & Cao

(2003), authors present a detailed study on the

mutual effects of different types of traffic (TCP and

UDP flows) over extremely low bandwidth WAN

links (128Kbps). In spite of the interesting results,

bitrates of nowadays user access lines and backbone

links are several times higher than the bitrates

considered in this work.

The design of protocols to deliver streaming

contents, maintaining fairness with other types of

services, is another interesting research field with a

large number of different approaches like Wu,

Claypool & Kinicki (2005), Song, Chung & Shin

(2002), Handley, Floyd, Pahdye & Widmer (2003)

or Balk, Gerla & Sanadidi (2003).

3 EXPERIMENTAL SETUP

With the goal of analyzing the influence of user

access lines variations, the particular architecture of

the service or the elements of the core network are

not representative. For this reason we have focused

the analysis on the emulation of the user access line.

To emulate different network conditions and user

access lines, we have combined Linux queuing

management (Iperf, n.d.) and a kernel module that

emulates packet loss and jitter, called NetEM (TC,

n.d.) as can be seen in Figure 1. We have used as

reference qualities for Downstream/Upstream

(US/DS) channels those offered by the Asturian

network operator Telecable: 640Kbps/128Kbps,

2Mbps/320Kbps and 4Mbps/640Kbps.

Figure 1: Model of the user access line implemented using

a linux router, 2FastEthernet cards and Netem.

To restrict the maximum bit rate that a user can

use to download/upload data, we have defined two

Token Buckets applied in the interfaces of the

router. This method of controlling bandwidth is

used, for example, in cable networks

(Laskshminarayanan and Padmanabhan, 2003).

Usually, network operators do not apply special

QoS policies for different types of user traffic,

except for VoIP telephony traffic. For this reason we

have taken the decision of deactivate the default

Linux Advanced Queuing Management (AQM).

Substituting AQM and including a FIFO (First In -

First Out) queue management. The size of the

buffers is set to a maximum latency of 1024ms, the

default on Cisco CMTS (Cable Modem Termination

System) (Kennedy and Atov, 2006).

We have also included NetEm module in the

DS/US to emulate some types of network delays like

Request-and-Grant control (Zhenglin, Chongyang,

2002). To know if with the current configuration of

network operator Telecable, the effect of Request-

and-Grant delays is significant we have tested a real

user connection. During the most loaded hours of the

day, RTT (Round Trip Times) between user and a

server installed close to the CMTS, was nearly

constant and equal to 69msec±3msec. Only 0.02%

of the RTT measurements present higher values.

Providing that the target of the study is the

influence of user access line constraints and

variations, we have deployed in the testbed a

simplified version of the network and the different

services that share the resources of the access line

(Figure 2).

FIFO Queue

Token Bucket

k packetsDS Kbps

TX Delay &

Packet loss

eth0

eth1

UPSTREAM CHANNEL

FIFO Queue

Token Bucket

n packets

US Kbps

Access Delay

& Packet loss

DOWNSTREAM CHANNEL

FIFO Queue

Token Bucket

k packetsDS Kbps

TX Delay &

Packet loss

eth0

eth1

UPSTREAM CHANNEL

FIFO Queue

Token Bucket

n packets

US Kbps

Access Delay

& Packet loss

DOWNSTREAM CHANNEL

PERFORMANCE OF AUDIO/VIDEO SERVICES ON CONSTRAINED VARIABLE USER ACCESS LINES

Linux Router

& Netem

Real Networks

Streaming

Server

Streaming

Client

Windows Media

Streaming

Server

Ftp & http

server

Http & Ftp

Client

Linux Router

& Netem

Real Networks

Streaming

Server

Streaming

Client

Windows Media

Streaming

Server

Ftp & http

server

Http & Ftp

Client

Figure 2: Testbed for the evaluation of streaming services

performance on constrained variable user access lines.

4 EXPERIMENTAL RESULTS

4.1 Single Streaming Session

We have tested the behavior of streaming sessions

when the channel capacity progressively decreases

from the maximum capacity of 4Mbps to 350Kbps,

and then, after reaching the minimum value of

350Kbps the quality was recovered upon reaching

4Mbps again. Intermediate values were 3.5Mbps,

3Mbps, 2.5Mbps, 2Mbps, 1.5Mbps, 1Mbps,

750Kbps, 500Kbps and 350Kbps. Contents were

produced using multi-rate encoding, with qualities of

700, 450, 225, 100, 50Kbps.

In the case of Windows Media delivery, the TCP

streaming connection consumes as much resources

as is possible (Figure 3). Unfortunately, using

windows streaming platform, each change in the

quality is performed stopping content reproduction

with the selected quality and starting a little later

with the new one. This can be due to the TCP API

that hides most network congestion indicators

(Chung and Claypool, 2006).

100 150 200 250 300 350 400

0.5

1.5

2.5

x 10

Time (seconds)

Delivery Rate (bps)

Real Networks

Windows Media

Figure 3: Delivery rate for Windows Media and Real

Networks when access line conditions change.

We have found that Real Networks streaming

sessions tend to react quickly to decrease in user

access lines. In spite of the use of Turboplay, the

delivery rate is not always equal to the available bit

rate of the line, for example, after setting 1Mbps as

the delivery rate, the delivery rate of Real Networks

was set to 700Kbps and the selected quality was

450Kbps. Once the channel conditions improved,

the quality of the content was not recovered at the

same speed. On most occasions, quality was

maintained at one of the lowest qualities until the

end of the session. When an adjustment of quality

occurs, this is done without stopping the playback of

contents, transparently for the client.

4.2 Streaming Session and FTP

To perform an analysis of the mutual influence

between TCP and UDP streaming traffic, we

perform some tests downloading a file using FTP

and after that, a streaming session is started. The bit

rate of the downstream channel progressively

decreased from 4Mbps to 200Kbps.

Results for Real Networks streaming are shown in

Figure 4. As can be seen, two different periods show

an aggressive behavior: the first 20 seconds of the

session and when channel capacity goes under

300Kbps. After that, there is a sudden decrease

down to 800Kbps. In the middle of these two values,

the bit rate achieved by the FTP session is almost

double that achieved by the streaming session.

Figure 4: Real Networks streaming session.

Initial delivery rates used by Real Networks

clients are not selected based on any measurement of

the actual situation of the line. This initial value

relies on the analysis of attainable delivery rate

performed during the installation phase. At the same

time, despite the fact that FTP obtains higher

bandwidths than the streaming session the

adaptation of streaming session delivery rate is

performed on extremely negative conditions. The

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

process of adjust the delivery rate seems to be

guided by the quality of the playback and the

encoding qualities of the stream.

In the case of Windows Media streaming (Figure

5) it can be seen that Windows Media streaming does

not share bandwidth with FTP for bit rates in the

access line higher than 2Mbps. When channel

conditions go below 300Kbps, the streaming session

slowly gets starved of bandwidth. This is possibly

due to a failure of the quality adjustment on

Windows Media platform that, as is stated

previously, reacts slowly to network congestion.

Approximately 30% of Windows Media streaming

sessions have achieved a maximum delivery rate of

500Kbps in presence of FTP competing traffic.

50 100 150

x 10

Time (seconds)

Bit Rate (bps)

Windows Media

FTP

Available delivery rate

Figure 5: Windows Media streaming session and FTP

delivery rate when bit rate decreases down to 350Kbps.

5 DELIVERY RATE ESTIMATION

FOR UDP STREAMING

Once it is clear that Real Networks streaming

sessions react slowly to improvements in the

delivery rate and that they do not consider the

congestion caused in the network, more advanced

techniques are needed to adapt the consumption of

streaming flows to the real conditions of the channel.

With this goal, we have analyzed the delivery

rate, packet loss, and inter arrival time of UDP flows

under different network conditions. Figure 6 shows

the CDF function of inter-arrival times for a 1Mbps

UDP over different access lines conditions. Results

are presented with and without simultaneous FTP

session. As can be seen, as the UDP flow

consumption grows variation of inter arrival time

gets bigger. Same results are obtained maintaining

the quality of the access line and varying the

delivery rate of the flow.

Figure 6: CDF of packet inter arrival time for different

access line conditions.

Based on this analysis, it is possible to add an

extra source of information to UDP streaming in

order to detect and react to variations in the user

access line quality. VTP (Balk, Gerla & Sanadidi

(2003) estimates the correct value for the delivery

rate based on the actual delivery rate measured on

the receiver’s side plus a smoothing function. The

approach presented in this paper is based on the

values of the variation of inter arrival times in the

receiver and inter departure times in the sender.

BART algorithm (Ekelin, Nilsson, Hartikainen,

Johnsson, Mangs, Melander and Bjorkman, 2006),

has previously used this type of estimation in order

to evaluate end to end available bandwidth on a

network path. Unlike BART, that generates extra

congestion on the network in order to estimate

available bandwidth in real time, the approach

presented in this paper relies on the data packets

generated by the streaming sessions.

Adding to sent packets the departure time from

the server, it is possible to obtain the difference

between departure times of consecutive packets. At

the same time it is possible to register arrival times

and calculate the inter arrival time. In this way, it is

possible to compare these two values and take a

decision about channel conditions. Using as a

measure of the congestion of the line the inter-

packet strain used in BART, it is possible to analyze

channel conditions and react correctly. Inter-packet

strain (ε) is defined as:

Being t

the inter arrival time of consecutive

packets and t

the inter departure time of the same

pair of packets. The formula shown before has been

modified from the original source in order to avoid

sudden increases due to packet compression; two

consecutive packets that arrive extremely close to

PERFORMANCE OF AUDIO/VIDEO SERVICES ON CONSTRAINED VARIABLE USER ACCESS LINES

the destiny. With inter-arrival times close to zero,

BART estimator presents great oscillations. For this

reason we have interchanged the position of t

and

From empirical observation (Figure 7) it is

possible to conclude that the average of inter-packet

strain increases as the delivery rate of the UDP flow

gets close to the available delivery rate. Whereas, if

the delivery rate increases over 80% of the delivery

rate of the access line, average value decreases as

long as the interference caused by the FTP session

decreases. Also, the variance of inter-packet strain

increases as the delivery rate of the UDP flow gets

close to the available delivery rate.

Figure 7: Inter-packet strain for a 1Mbps UDP flow and

different DS rates (Mbps).

5.1 Experimental Results

We have tested the behavior of inter-packet strain as

an estimator of channel congestion. We have

deployed a simplified version of a streaming server

and client that establish a TCP control connection.

After a short negotiation, the server starts sending

UDP packets at a predefined rate to the client. These

packets include the departure time. The client

estimates channel conditions based on inter-packet

strain and takes a decision about the suitability of a

delivery rate adaptation. Delivery rate adaptation

messages are sent from the client to the server using

the TCP control connection in the same fashion as

Real Networks streaming.

In this first approach only fixed packet sizes are

considered and the thresholds to cause a delivery

rate adaptation were determined experimentally

during a set of trials. To avoid the effects of

temporary drops in network conditions, the mean

value and the variance of packet strain is calculated

every 0.2 seconds. Possible delivery rate values were

selected at 200Kbps steps, ranging from 200Kbps to

2Mbps, simulating a multiple-bit rate file. The value

of the delivery rate, used as the base in the

estimations, is obtained from the size of the packets

divided by the difference between the departure

times of two consecutive packets from the server.

Results, when available delivery rate varies, are

shown in Figure 8. Using inter-packet strain in

heavy loaded user access lines we have achieved

values for the delivery rate that accomplish these

basic properties:

• During phases of channel stability no adaptations

are generated.

• When congestion increases, the delivery rate is

adjusted, once or more, until the estimator shows

no congestion. These adaptations are non erratic,

following the same pattern (increase or decrease)

as the conditions of the access line.

• The estimator behaves in a conservative way,

maintaining UDP flow consumption to values

close to one half of the attainable delivery rate or

lower.

Figure 8: UDP streaming based on packet strain delivery

rate estimation.

At the same time, it can be seen in Figure 8 that

the FTP session never gets starved of bandwidth.

Also, reaction to the increase of capacity is much

faster than that obtained using Real Networks and

follows the evolution of the access line, sharing the

available delivery rate with the FTP session.

6 CONCLUSIONS

Previous studies have shown that UDP streaming is

a bandwidth intensive application that does not

perform a fair share of resources with TCP

connections. As is described in this paper, the

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

Turboplay feature presents strong unfriendly

behavior with other flows at the beginning of the

session, especially when estimations of user access

line delivery rate are aborted. Adaptability of Real

Networks streaming sessions seems to be completely

driven by the set of encoding qualities and the

quality of the playback, without extra measurements

of channel congestion.

To solve these problems, some techniques, like

inter-packet strain measurement, could be useful to

detect congestion and react so reducing it. In this

way, the interference between streaming sessions

and other services could be minimized. Combining

session quality measurements (like percentage of

packet loss and client buffer size) with congestion

estimators is possible to improve the behavior of

UDP streams.

7 FUTURE WORK

Streaming platforms typically use variable packet

sizes and, some of them, even variable bit rates. It is

necessary to study the reaction of inter-packet strain

in the presence of variable packet sizes and variable

inter arrival times. Also, the reaction of inter-packet

strain when resources are shared with another type

of services has to be carefully evaluated.

It is also necessary to perform an in-depth study

of the suitability of inter-packet strain estimator

when the packets generated in the server have to

travel through several routers, interacting with a

variable number of flows.

Finally, the integration of inter-packet strain on a

real streaming platform is a very interesting field of

study.

ACKNOWLEDGEMENTS

This work was supported in part by Telecable, La

Nueva España and Asturies.com within the projects

of NuevaMedia, Telemedia, ModelMedia and

MediaXXI and the Spanish National Research

Program within the project INTEGRAMEDIA

(TSI2004-00979).

REFERENCES

Li, M., Claypool, M., Kinicki, R. (2002), MediaPlayer

versus Real Player – A Comparison of Network

Turbulence, Paper presented at ACM SIGCOMM,

Marseille, France, 2002.

Kuang, T., Carey, W. (2002), A measurement study of

Real Media Streaming Traffic, Paper presented at

ITCOM, Boston, Massachusetts USA, 2002

Chung, J., Claypool, M. (2006), Empirical Evaluation of

the Congestion Responsiveness of Real Player Video

Streams, Kluwer Multimedia Tools and Applications,

Volume 31, Number 2, November 2006.

Real Networks, (n.d), Retrieved March 12

, 2007, from

Boyden, S., Mahanti, A., Williamsom, C. (2005),

Characterizing the Behaviour of RealVideo Streams,

Paper presented at SCS, Philadelphia, USA, 2005.

Doshi, R., Cao, P. (2003), Streaming traffic fairness over

low bandwidth WAN links, Paper presented at WIAPP

2003. San Jose, CA, USA, 2003.

Wu, H., Claypool, M., Kinicki, R. (2005), Adjusting

Forward Error Correction with Temporal Scaling for

TCP-Friendly Streaming MPEG, ACM Transactions

on Multimedia Computing, Communications and

Applications, Vol 1, Issue 4, November 2005.

Song. B., Chung, K., Shin, Y. (2002), SRTP: TCP-

Friendly Congestion Control for Multimedia

Streaming, Paper presented at ICOIN, Korea, 2002.

Handley, M., Floyd, S., Pahdye, J., Widmer, J., TCP

Friendly Rate Control (TFRC): Protocol Specification,

RFC 3448, January 2003.

Iperf: Testing the limits of your Network. TCP/UDP

Bandwidth Management Tool, (n.d) Retrieved March

, 2007 from http://dast.nlanr.net/Projects/Iperf/

TC. Linux QoS Control Tool, (n.d), Retrieved March 12

2007 from http://www.rns-nis.co.yu/~mps/

NetEM. Network Emulation (n.d), Retrieved March 12

2007 from http://linux-net.osdl.org/index.php/Netem

Laskshminarayanan, K., Padmanabhan, V. (2003), Some

findings on the Network Performance of Broadband

Hosts, Paper presented at ACM conference on Internet

measurement, Florida, USA, 2003.

Kennedy D., and Atov, I., (2006) Implementing a Testbed

for the Evaluation of FAST TCP in DOCSIS-based

Access Networks, Technical Report 060119A,

Swinburne University of Technology, 2006.

Zhenglin, L., Chongyang, X., (2002), An Analytical

Model for the Performance of the DOCSIS CATV

Network, The Computer Journal, Vol 45, No. 3, 2002.

Ekelin, S., Nilsson, M., Hartikainen, E., Johnsson, A.,

Mangs, J.-E., Melander, B., Bjorkman, M., (2006),

Real-Time Measurement of End-to-End Available

Bandwidth using Kalman Filtering, 10th IEEE/IFIP

NOM 2006, Vancouver, Canada.

Balk, A., Gerla, M., Sanadidi, M., (2003) Adaptive

MPEG-4 Video Streaming with Bandwidth

Estimation, Lecture Notes In Computer Science; Vol.

2601, 2003.

PERFORMANCE OF AUDIO/VIDEO SERVICES ON CONSTRAINED VARIABLE USER ACCESS LINES