WEB TRAFFIC ACCELERATION OVER CELLULAR

NETWORKS BY USING A COMPRESSING PROXY SERVER

Yair Toaff and Ariel J. Frank

Department of Computer Science, Bar-Ilan University, Israel

Keywords: Internet, cellular, proxy, compression, Performance Enhancing Proxy (PEP), GSM, HSCSD

Abstract: Many Web clients today are connected to the Internet via low speed computer links such as cellular

connections. In order to efficiently use the cellular connection for Web access, the connection must be

accelerated using a Performance Enhancing Proxy (PEP) as a gateway to the Web. In this paper we

investigate the challenges created by the use of PEP. In order to mitigate the performance bottlenecks, we

enhanced the PEP by adding to it a “pre-getting” ability. We tested the enhanced PEP over GSM and

HSCSD cellular networks, as well as using a cellular network software simulator. Our experiments with

enhanced PEP achieved the following results: 1) Average improvement of about 60% over the HSCSD

network throughput can be achieved, 2) The Web page structure has a significant impact on the resulting

performance. In conclusion, using an enhanced PEP can make the experience of browsing the Web over

cellular networks much faster and less frustrating.

1 INTRODUCTION

Cellular access to the Internet began about a decade

ago. The general characteristics of data transmission

over cellular networks are a high bit error rate when

compared to regular phone lines, frequent

disconnects, and a generally slow transfer rate when

compared to other means of data transmission. For

these reasons, effective Internet browsing over

cellular networks requires some means of

acceleration. One such mean is the use of a

compressing proxy server. The compressing proxy

server is the focus of this paper.

We will first review the two types of technical

challenges that we wish to analyze and resolve.

The first challenge is related to Internet browsing

over cellular networks in general. The component

issues are the time required to create a TCP

connection (RFC1072 1988, Stevens 1996), and the

synchronous protocol that the browser implements

(Toaff 2002). Looking at the graph of the bandwidth

used while downloading a Web page over a GSM

(Global System for Mobile Communications)

connection (Scourias 1996), we can see (Figure 1)

that the bandwidth is not fully utilized.

Figure 1: Web page download bandwidth graph

A partial solution for both these challenges is the use

of a Performance Enhancing Proxy (PEP). PEP

servers can usually provide a solution for the low

bandwidth of cellular networks by using

compression. Both client and server compress the

data transferred over the cellular link. The data is

then uncompressed at the target (Liljeberg 1996,

Brown and Singh 1997, Toaff 2002).

The second set of issues is related to the use of PEP.

When using PEP, the average actual bandwidth used

drops significantly. It would be expected that if the

compression ratio is about 1:3 – the resulting time

ratio would be similar. What we have found,

however, is that the average actual bandwidth used

drops but the time ratio obtained is only about 3:4.

This phenomenon is the result of several reasons.

The time needed for a request to reach the server

remains the same but compression times must be

added, causing a delay. There is less data to transfer

Yamada H. and Kawaguchi A. (2004).

A METHOD FOR THE PERFORMANCE ANALYSIS OF INTEGRATED APPLICATION SERVICES - Simulating the execution of integrated application

services coordinated by workﬂow-driven broker-servers.

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 40-45

DOI: 10.5220/0001382900400045

 SciTePress

so the periods when data is downloaded are shorter.

In the experiments described in section 5.1, although

the size of the downloaded data is about 30% of the

original data (120KB instead of 379KB), the

download time drops only to about 50% of the

original time (160s instead of 345s). This result still

indicates inefficient use of the available bandwidth.

In the advanced cellular networks, such as HSCSD

(High Speed Circuit Switch Data) (ETSI 1997), the

problem appears to be more severe (data ratio about

1:3, time ratio about 2:3) due to the asymmetric

nature of the network (Toaff 2002).

The paper is organized as follows:

 Section 2 reviews related research works.

 Section 3 introduces the logical model for the

pre-getting PEP.

 Section 4 describes the environment used for

the experiments (both software and

hardware).

 Section 5 describes the experiments that

were carried out and the conclusions reached

from them.

 Section 6 finally concludes with a discussion

and suggests additional issues that may

require further investigation.

2 RELATED RESEARCH WORKS

In this section we briefly review previous research

works in related topics.

These related topics are:

 Data communication over cellular networks.

 Internet browsing over cellular networks.

2.1 Data communication over cellular

networks

Several research works (Xylomenos and Polyzos

1999, Brown and Singh 1997, Vardalis 2001,

Balakrishnan 1997, Tsaoussidis and Matta 2002)

discuss the problems with data communication over

cellular networks and suggest several methods of

improving the existing TCP protocol. An important

idea mentioned in these papers is data compression

over the cellular connection but the idea itself was

not implemented and tested, and the challenges

caused by the use of compression were not

discussed.

2.2 Internet browsing over cellular

networks

Article (Liljeberg 1996) suggests a client-server

based proxy that handles the pre-getting of page

embedded objects. This idea is similar to our work;

however, content compression was not implemented.

The paper simply states that the results of

compression are easily predicted. But as implied

above, the results may not be as clear-cut as that

paper predicts.

The main problem we have found with the described

work is that it ignores new asymmetric cellular

networks architectures (such as HSCSD) and

assumes a network supplying stable bandwidth and

equal bit rate for both upstream and downstream

data transfer. In the HSCSD network we used, the

downstream bit rate is 3 times the upstream bit rate.

This influences performance since the time required

for the request to reach the server remains the same,

but the time required for the response to reach the

client decreases by a factor of 3 (in HSCSD

compared to GSM). This means that the full

utilization of the downstream channel will require

that 3 times the number of requests be sent for the

same bandwidth. This network asymmetry causes

the aforesaid problems to become more significant.

Due to these problems, pre-getting is needed to

enhance the utilization of the downstream channel

and compression must be applied to both channels.

3 LOGICAL MODEL FOR

ENHANCED PEP

The main object of our research has been the

enhancement of an existing PEP with the purpose of

resolving the problems previously encountered. We

have found that adding a pre-getting mechanism to

PEP achieves this goal.

In the pre-getting method, no speculative algorithms

are required. Instead, pre-getting evaluates and

modifies the way the browser requests information

from Web servers and how the browser validates

that information.

The pre-getting method works as follows:

1. The proxy client parses the HTML page,

obtains a list of the embedded objects,

compresses it and transmits the list of required

objects to the server.

2. The proxy server builds the HTTP requests

from the list, retrieves and compresses them,

and returns them to the client.

3. The proxy client decompresses the

responses and caches them.

WEB TRAFFIC ACCELERATION OVER CELLULAR NETWORKS BY USING A COMPRESSING PROXY SERVER

4. When the browser requests the elements,

they are ready for delivery, stored in the proxy

client’s cache.

The result of using pre-getting is that all the

embedded objects from a certain HTML page are

downloaded without time gaps. The idle download

time previously observed is eliminated. The

observable result is that the browser finishes

downloading the page faster.

The parameter we have used for measuring the

resulting improvement is the time required to

download a Web page. We assume that the user is

willing to “pay” for the reduced download time with

some loss in the quality of the pictures contained in

the page.

As the compression ratio can be assumed to be

constant, therefore the improvement can be

measured by the time required for the Web page to

download. Our goal in enhancing the PEP is for the

achieved time ratio to be similar to that of the

compression ratio.

4 EXPERIMENT ENVIRONMENT

This section describes the environment we have used

for our experiments. The description includes the

software used in the experiments, the various

components, and the hardware infrastructure.

4.1 Software

The general characteristics of the system are:

 Use of a client/server system.

 The client is connected to the server by a

cellular connection.

 Both client and server support compression.

The commercial software used was the Apache

Web Server and MS Internet Explorer browser.

Client

The regular compressing proxy client is

composed of a communications module, a

compression module, and a multiplexor module in

charge of multiplexing the multiple browser

connections into a single server connection.

To the enhanced compressing proxy client we

added an HTML parser module and a cache module.

Server

The regular compressing proxy server is

composed of a communications module, a

compression module, and a multiplexer module. To

the enhanced compressing proxy server we added a

URL handler module for transforming the URLs into

full HTTP requests.

4.2 Cellular network simulator

A software simulation of a cellular link was used

due to the high cost of airtime on cellular networks

and the requirement of the cellular link to supply

constant bandwidth thus enabling the results to be

compared.

4.3 Hardware infrastructure

Two MS Windows based workstations were used for

the experiments. The communication hardware for

the experiments on the real cellular network was a

PCMCIA Nokia card phone. This card phone

supports both GSM and HSCSD network

connections.

For the simulator experiments, a 100Mb/s LAN

was used to connect the computers. The

experimental configuration is described in Figure 2.

Figure 2: Experimental configuration overview

5 EXPERIMENTS AND

CONCLUSIONS

This section describes the experiments. The first set

of experiments analyzes performance over real

cellular networks (GSM and HSCSD). The second

set of experiments used the cellular network

simulator. Following the description of each

experiment, the resulting conclusions are discussed.

5.1 Real cellular network experiments

Our experiments using the pre-getting method

were carried out as follows. A variety of Internet

sites were copied onto a local Web server (to

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

eliminate delays caused by communication with the

content providers). Each of the chosen sites was

downloaded three times:

• Without use of a proxy.

• Using the regular PEP.

• Using the pre-getting PEP.

For each download, 3 parameters were recorded:

1. The total amount of data.

2. The time required for download.

3. The average transfer rate.

After each download, the browser’s caches were

flushed (so the data and time totals could be

compared).

Conclusion 1. The pre-getting method improves

the compressing proxy bandwidth usage by up to

90% of the measured bandwidth with no

compression used.

As can be seen in Table 1, the pre-getting

method significantly improves the utilization of the

available bandwidth.

Table 1: Cellular downstream average transfer rate (in

bps)

Cellular transfer rate

Methods

GSM HSCSD

Theoretical

9600 43200

No compression 8765 18595

Regular PEP 6062 9830

Pre-getting PEP 7946 17121

On the GSM network, the used bandwidth

dropped by about 30% when using a regular PEP.

The use of the pre-getting PEP lowered the drop to

just under 10%.

On the HSCSD network, the used bandwidth

dropped to about 50% when using a regular PEP.

The use of pre-getting PEP lowered the drop also to

just under 10%.

The improvement in the used bandwidth brings

the resulting time ratio closer to the compression

ratio. The 10% drop in the bandwidth utilization is

due to the data passing more stages and to the

computations requiring additional time (parsing,

(de)compression).

5.2 Cellular network simulator

experiments

The experiments of the pre-getting methods in the

cellular network simulator were done in a similar

manner, except that in this case the simulator

replaces the cellular network

5.2.1 Imitating the cellular network

The first phase of the simulator experiments was to

verify that the simulation closely simulates real

network conditions, enabling reliable use of the

results. This was tested using the parameters of the

HSCSD network.

The parameter that was changed between the

experiments is protocol overhead. The overheads

that were tested included 148 bytes, 248 bytes and

348 bytes per frame.

Conclusion 2. The simulator imitates the

behavior of the real cellular network with good

precision.

It appears that simulated results based on a 23%

loss in the bandwidth utilization (overhead of 348

bytes per frame), result in an effective bit rate that is

almost identical to the HSCSD experiments.

5.2.2 Web page structure influence on the

performance

Analyzing the structure of the Web pages used we

identified that the number and size of embedded

objects in a page affects the download time. In sites

that contain a small number of pictures, (i.e.,

http://www.google.com – 2 small pictures), the time

difference between using a regular PEP and a pre-

getting PEP is very small.

On the other hand, there are sites that contain a

large number of small pictures, (i.e.,

http://www.barnesandnoble.com – when used had 51

pictures, 39 of them with sizes below 1K). On these

sites, the improvements obtained by utilizing a pre-

getting method are very significant. The resulting

download time dropped to about 30% of the

download time using a regular PEP.

The average improvement on all the Web pages

on a local server was about 50%. The bandwidth

utilization ratio on the “small” sites was about 1.3

(i.e., the bandwidth used without PEP was 1.3 times

higher than with pre-getting PEP), while on the

“large” sites the ratio was about 0.7. The following 2

experiments analyzed the relationship between Web

page structure and pre-getting performance, by using

different Web pages built specially for this purpose.

Experiment 1 description

To further analyze the relationship between Web

page structure and performance we added two

additional pages that represent the extreme cases.

The first page contains 6 pictures larger than 100KB

and 8 pictures larger than 50KB. The second page

WEB TRAFFIC ACCELERATION OVER CELLULAR NETWORKS BY USING A COMPRESSING PROXY SERVER

contains 24 pictures smaller than 1KB and 28

pictures smaller than 5KB. We will refer to these 2

pages as big and small, respectively.

Conclusion 3. The performance of the pre-

getting method depends on the Web page

structure.

Table 2: Experiment results on “extreme” cases

Big Small

Time (s) 306.6 83.6

Total data 1252 163.3

Compression

Average KB/s 32.6 15.6

Time (s) 158.7 64.9

Total data 373.3 73.4

Compressing

proxy

Average KB/s 18.8 9.05

Time (s) 90.8 20.2

Total data 373.3 73.4

Pre-getting

proxy

Average KB/s 32.9 29.1

The experiment results on these pages were as

expected (see Table 2). As can be seen, the

bandwidth utilized while using the pre-getting PEP

remains stable around 30KB/s (on both types of

pages). The bandwidth, while not using PEP at all, is

affected by the page structure and drops by about

50% and while using the PEP it is about 60% of the

bandwidth without a proxy.

Experiment 2 description

The previous experiment tested extreme cases. In

order to analyze cases between the two extremes we

performed an additional experiment. For this

experiment, we collected a series of 8 pictures with

sizes from 43 bytes to 80KB. For each of these

pictures, we built a Web page that contained the

picture 15 times.

The effect of the picture sizes on the bandwidth

utilization can be seen in Figure 3. The basic

conclusion that a large amount of small pictures

degrades bandwidth utilization still holds.

Bandwidth utilization is in direct relationship to the

pictures sizes. The only exception is in the case that

the picture is compressed extremely well (from

30,567B to less than 3,000B). Even in these cases,

the bandwidth utilization obtained by using the pre-

getting PEP is higher than the bandwidth utilization

obtained by using a regular PEP. In normal cases

(compression rate of 50% to 80%), we have found

that the bandwidth utilization is similar to bandwidth

utilization without a PEP and even exceeds it.

Figure 3: Bandwidth utilization in relation to pictures sizes

Only in the most extreme case (the smallest

pictures) a drop in the used bandwidth was

identified. But even with this drop, bandwidth

utilization obtained by using other methods is

exceeded.

Conclusion 4. A pre-getting compressing proxy

gives us the ability to download large Web pages.

When attempting to download a large Web page

using the cellular network a timeout on the browser

was obtained. The browser attempts to download the

page for 5 minutes and if the download is not

completed the browser returns an error message. If

we calculate the page size that will result in a

timeout, we get 1.5MB (43,200/8 * 60 * 5 =

1,620,000) on a HSCSD network.

If we assume a compression ratio of 1:4 and

bandwidth utilization that drops to about half, we

achieve the ability to download Web pages that are

twice as large without getting a timeout. When we

use the pre-getting PEP (so the bandwidth used is

about the same as without a proxy) we can download

Web pages that are four times as large (i.e., up to

6MB in size).

6 DISCUSSION

The work presented here investigates the problems

related to use of a compressing proxy server over

cellular networks. The main problem was the

effectiveness of bandwidth use. We saw the drop in

bandwidth while using the PEP and studied its

sources.

The next step was to offer a solution to this

problem by adding a pre-getting mechanism to the

compressing proxy server. We implemented this

feature into the PEP and tested the effectiveness of

the solution.

We saw that the pre-getting method boosted the

performance of the PEP. With an average

compression ratio of about 65% we obtained an

improvement of about 60% in the download time,

compared to 25% while using the regular PEP.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

When we reviewed the results, we saw that the

bandwidth ratio while using the pre-getting proxy

server varied from 0.5 to 1.5. In order to identify the

source of this inconsistency in a sterile environment,

we built a cellular network software simulator.

Our experiments indicated that the structure of

the Web page was the root cause of the variance in

bandwidth utilization. We discovered that the best

performance is on pages with a large number of

embedded small pictures and that the worst

performance is with pages holding a small number

of large pictures. We constructed several sample

pages to analyze the relationship between the sizes

of the pictures embedded in the Web page to the pre-

getting performance.

In conclusion, we have identified the problems

which occur while using a regular compressing PEP,

and introduced a mechanism that enhances an

existing PEP in a way that enables improved Web

surfing over a cellular network. Without this

enhancement, Web surfing is slow, there is no direct

relationship between quality loss and improvements

in download time and it is practically impossible to

download large Web pages. With this enhancement,

the ratio between the loss in quality (the

compression ratio) and the gain in time is close. In

other words, if the user is willing to lose 90% of the

data by reducing the quality of the pictures in the

Web page, the page will be gotten 10 times faster.

Future research topics concerning the pre-getting

method are:

• Adding compression

for various data types –

in this work, only GIF and JPG graphic files

were explicitly compressed and at a constant

ratio. All other types of files were

compressed with the standard zlib

compression.

• Improving the communication protocol

–

The communication protocol is

implemented over TCP/IP. It can be

improved by using the protocol

improvements referred to in subsection 2.1.

REFERENCES

Balakrishnan, H., Padmanabhan, V. N., Seshan, S., and

Katz, R. H., 1997. “A Comparison of Mechanisms for

Improving TCP Performance over Wireless Links”,

IEEE\ACM Transactions on Networking, 5(6), 756-

769.

Brown, K., and Singh, S., 1997. “M-TCP: TCP for Mobile

Cellular Networks”, ACM Computer Communication

Review, 27(5), 19-43.

ETSI 1997. “High Speed Circuit Switch Data (HSCSD) –

stage 1”, European Telecommunications Standards

Institute, GSM specifications 02.34.

Liljeberg, M., Helin, H., Kojo, M., and Raatikainen, K.,

1996. “Enhanced Services for World-Wide Web in

Mobile WAN Environment”, CS Technical Report

No. C-1996-28, University of Helsinki.

RFC1072, 1988. “TCP Extensions for Long-delay Paths”,

http://www.cis.ohio-state.edu/Services/rfc/rfc-

text/rfc1072.txt

Scourias, J., 1996. “Overview of the Global System for

Mobile Communications”,

http://ccnga.uwaterloo.ca/~jscouria/GSM/gsmreport.ht

ml.

Stevens, W. R., 1996. “TCP/IP Illustrated”, Vol. 1,

Addison-Wesley, Chapters 18, 19, 21.

Toaff, Y., 2002. “Web Traffic Acceleration over Cellular

Networks by using a Compressing Proxy Server”,

Master Thesis, Dept. of Computer Science, Bar-Ilan

University, Ramat-Gan.

Tsaoussidis, V., Matta, I., 2002. “Open issues on TCP for

Mobile Computing”, Journal of Wireless

Communications and Mobile Computing, Wiley

Academic Publishers, 2(2),

http://www.cs.bu.edu/techreports/pdf/2001-013-open-

issues-tcp-wireless.pdf.

Vardalis, D., 2001. “On the Efficiency and Fairness of

TCP over Wired/Wireless networks”, Master Thesis,

SUNY at Stony Brook, New York.

Xylomenos, G., and Polyzos, G., 1999. “TCP and UDP

Performance over a Wireless LAN”, Proc. IEEE

INFOCOM.

WEB TRAFFIC ACCELERATION OVER CELLULAR NETWORKS BY USING A COMPRESSING PROXY SERVER