PERFORMANCE EVALUATION OF TCP ALGORITHMS ON

HYBRID WIRED/WIRELESS LAN TEST-BED

Apostolos Kotsiolis, Christos Antonopoulos and Stavros Koubias

Applied Electronics Laboratory, Department of Electrical & Computer Engineering

University of Patras, Rio Campus, Patras, Greece

Keywords: Transport Layer, TCP Algorithms Evaluation, WAN Test-bed, Linux, Congestion and Admission Control,

Wireless/Wireline Networks.

Abstract: This paper aims to provide comparative performance evaluation of various available TCP algorithms over a

hybrid local area network comprising by, both wired and wireless sections. Although TCP originally was

designed aiming to control internet traffic, LAN networks increasingly depend on new TCP protocol

versions to provide congestion control, fairness and optimum resource utilization. Additionally wireless

LANs’ popularity increase exponentially. Based on the aforementioned observations, the objective of this

paper is two-fold. Firstly, to present an adequate test-bed enabling, through a significant number of

experiments, an accurate performance evaluation of five different TCP versions available by a typical Linux

distribution. Secondly based on the conclusions extracted, propose efficient, yet low complexity

modifications, able to improve network performance of the considered scenarios.

1 INTRODUCTION

Transmission Control Protocol (TCP) was designed

to provide a reliable, connection-oriented, transport

service over the unreliable service provided by the

Internet Protocol (IP). The effect of the TCP/IP

protocol combination was so profound that formed

the backbone of the existing wired computer

networks with the Internet being the most

characteristic example.

During the last decade there has been a rapid

increase on the number of Local Area Networks

(LANs) implementations covering a large range of

end user demands. Furthermore, the most significant

advance of networking during the last years is

expressed by wireless connectivity providing a

whole new area of networks that lead to an even

bigger expansion of LANs’ size and popularity

(Mohaparta, 2005). However, on the other hand the

nature of wireless connection has been the main

reason for TCP congestion control suboptimal

behaviour in WLANs, as TCP is not well suited to

error prone links (Mascolo, 2001).

These reasons lead to the evolution of TCP

protocol family which had to maintain its significant

role as the robust and reliable Transport Layer

Protocol in addition to providing solutions to the

congestion control problems arisen by the growing

complexity of LANs and the Internet.(Mascolo,

2001), (Grieco, 2004).

The main objective of this paper, is

comparatively evaluating five of the most popular

TCP congestion control algorithms that are provided

by the Linux kernel, (BIC, New Reno, H-TCP,

Vegas and Westwood+) on a typical network

environment. A critical difference of the presented

approach, as opposed to many already presented

research efforts focusing on simulation based

evaluation, is that the experiments were carried out

on a real local wired/wireless network. Additionally

measurements were extracted directly from the

Linux kernel as well as by using well known

network analyzing applications, for verification as

well as system level measurements, as opposed to

relying solely to application level measurements.

Our goal is to evaluate the performance and

behaviour of these algorithms that are designed

aiming to control network traffic on a number of

large data transferring scenarios taking place on a

hybrid wired/wireless LAN test-bed.

Additionally, based on these measurements,

modifications are proposed concerning a specific

algorithm evaluated. Through these measurements,

Kotsiolis A., Antonopoulos C. and Koubias S. (2010).

PERFORMANCE EVALUATION OF TCP ALGORITHMS ON HYBRID WIRED/WIRELESS LAN TEST-BED.

In Proceedings of the International Conference on Data Communication Networking and Optical Communication Systems, pages 73-79

DOI: 10.5220/0002981900730079

 SciTePress

potentially significant benefits are identified and

presented.

The rest of this paper is structured as follows:

Section 2 presents the TCP versions under

evaluation. In section 3 the test-bed setup is

described followed by the performance analysis in

section 4. Section 5 includes the presentation and

evaluation analysis of the proposed modifications

while in section 6 the main conclusions are

summarized.

2 PROMINENT TCP

PROTOCOLS

Five congestion control algorithms are evaluated

through numerous performance measurements

including, BIC, New Reno, H-TCP, Vegas and

Westwood+. The following brief description aims to

reveal each one’s philosophy and specialized

mechanisms.

Binary Increase Control TCP (BIC) (Xur, 2004)

is designed focusing on fast long-distance networks

and aims to satisfy three main performance criteria:

RTT fairness, TCP friendliness and scalability.

New Reno TCP (Floyd, 1999) is a modification

of the classic Tahoe/Reno algorithm that improves

retransmission during the fast recovery phase of

TCP Reno.

H-TCP (Leith, 2004) is a congestion control

algorithm presented by the Hamilton Institute

suitable for deployment in high-speed and long-

distance networks.

Vegas TCP (Brakmo, 1995) was the first attempt

to depart from the loss-driven paradigm of the TCP

by introducing a mechanism of congestion detection

before packet losses.

Westwood+ TCP algorithm (Mascolo, 2001),

(Grieco, 2004) is based on end-to-end estimation of

the bandwidth availability along the TCP connection

path.

Due to the specific features and characteristics of

the aforementioned TCP protocols are considered in

the following evaluation process.

3 TEST-BED SETUP &

CONFIGURATION

As depicted in “Fig. 1” two different topologies

comprise the evaluation test-bed. The goal of the

wired topology is to evaluate network performance

in an optimal case were server-client connection is

interrupted only by a switch and therefore comprises

by two wired sections supporting 100Mbps

connections. Contrary, the objective of the hybrid

topology is to stress the TCP operation by inserting a

wireless section theoretically providing 54Mbps

communication. To implement this topology, data

from the switch are now forwarded to a router with

wireless capabilities and then to the final destination

(and vice versa of course). Additionally between the

wireless router and the computer with wireless

interface, walls intervene in order to emulate a

suboptimal wireless connection.

In all cases communicating computers are

running a 2.6.14.3 based Linux modified,

recompiled kernel. The kernel, depending on the

experiment scenario, was configured to support each

TCP version presented in section 2. There was an

FTP server daemon running on the “Server” station,

while a specific 800Mbyte test file used is

transferred during all experiments. The selected file

assures sufficient transferring period in order to

expose the behavioural and performance

characteristics of the TCP algorithm under

evaluation.

Measurements are taken using tcpdump, tcptrace

and xplot together with the kernel extracted

measurements using the

printk function. In order

to comparatively evaluate download (server-to-

client) and upload (client-to-server) directions, five

types of transferring scenarios were considered,

including one or two simultaneous data flows,

• One download/ upload

• Two parallel downloads/ uploads

• A two data flows crossed download/upload.

From a hardware point of view all components

supported adequate communication rates, so as not

to impose bottlenecks in the experiments. Finally it

is noted that all measurements have been taken from

the “Client” station.

Figure 1: Test-Bed.

DCNET 2010 - International Conference on Data Communication Networking

4 COMPARATIVE

PERFORMACE EVALUATION

In order to clearly and objectively investigate the

influence of each different factor under evaluation,

three different parameters will be used as means of

comparison. Each one of them focuses on a different

area of influence such as, network topology, data

transfer scenario and congestion control algorithm.

4.1 The Effect of Network Topology

As indicated by respective measurements, the

differences of the two network topologies have an

immediate effect on the behaviour of cwnd and RTT.

However it is observed that for the same topology

the TCP algorithm version it is not a critical factor.

Thus indicative measurements concerning the H-

TCP algorithm will be presented since analogous

conclusions are drawn from all cases regarding

different algorithm.

Thus, in “Fig. 3” it is depicted that in wired

scenario the range of cwnd value’s fluctuation is

quite limited indicating a steady behaviour with an

average value of 18140 bytes. On the contrary for

wireless scenario in “Fig. 4” a rather unstable

behaviour is depicted with abrupt and frequent

fluctuations reaching about 50% of the actual

window value, while the mean value is significantly

higher reaching about 32162 bytes. Round trip time

(RTT) measurements also provide interesting

insights. In “Fig. 2” the grey graph clearly indicate

the effect of packet loss of the wireless link in

hybrid topology which triggers the end of congestion

avoidance phase and results the decrease of cwnd.

As observed in the hybrid topology the experiment

required almost twice the time period due to an RTT

overhead of 20-30msec. In addition, as shown, there

are short periods of time (spikes) where RTT in

hybrid topology exceeded even 100msec severally

degrading network performance.

The result of this performance degradation is clearly

shown on the respective throughput graphs in “Fig.

5”. The combination of a higher average RTT of

50msec and the periods of 100+msec together with

the wider range of cwnd in “Fig. 4” lead to a 45%

higher achieved throughput for wired topology and

consequently to an almost 50% delay increase for

the hybrid topology.

Figure 2: RTT of “H-TCP 1 Upload Scenario” on both

topologies.

Figure 3: cwnd of “H-TCP 1 Upload” Scenario on the

Wired topology.

Figure 4: cwnd of “H-TCP 1 Upload” Scenario on the

Hybrid topology.

PERFORMANCE EVALUATION OF TCP ALGORITHMS ON HYBRID WIRED/WIRELESS LAN TEST-BED

Figure 5: Throughput of “H-TCP 1 Upload” Scenario on

both topologies.

Figure 6: Throughput of “Crossed” Scenario on Wired

topology.

Figure 7: Throughput of “Crossed” Scenario on Hybrid

topology.

4.2 The Effect of Data Flow Scenario

The number and direction of the coexisting data

flows (either from the server towards the client and

vice versa) in the channel clearly affect the

efficiency of TCP congestion control mechanism as

well. This observation is clearly depicted in Table 1

presenting the time durations required for the file

transfer completion on the hybrid topology

scenarios. There is a notable difference between the

upload and download scenarios since the upload

flows complete their transfer sooner. Furthermore

the significance of the direction is also evident

through the bandwidth distribution measurements

since, the upload data flows capture a range of

60% up to 80% of the available bandwidth while

the download flows only from 20% up to 40%.

Analysing the measurements in more detail it is

observed that in the one flow scenario, upload is

approximately 30% faster than download where on

the crossed scenario the upload flow is 40% faster

than the download flow. The same observation can

be extracted from “Fig 7” concerning hybrid

topological scenarios. Therefore, during an FTP file

transfer where two crossed flows coexist, TCP tends

to favour the client-to-server flow at the expense of

the server-to-client flow leading to faster file

transfers for the upload direction. Furthermore,

throughput graphs of scenarios concerning two

crossed flows comparative analysis between hybrid

and wired topology offer additional valuable insights

as shown in “Fig. 6” and “Fig. 7”. Focusing on the

wired topology the two crossed flows converge

rapidly to a fair distribution of the channel’s

bandwidth. On the other hand such a convergence

was not observed in the hybrid topology;

significantly degrading download direction

performance since the respective throughput graph

exhibits increase tendencies only after the upload

session is concluded.

In the two parallel flow scenarios (either

download or upload) TCP is able to guarantee an

equal 50% share of the channel’s bandwidth to the

two competing flows. This behaviour was confirmed

for all the tested algorithms and both network

topologies.

4.3 The Effect of Configured TCP

Algorithm Version

Using the measured transfer time durations from

Table 1 together with the throughput graphs of “Fig.

7”, we will try to compare the five algorithms

DCNET 2010 - International Conference on Data Communication Networking

Table 1: Hybrid Topology Transfer Time (minutes).

Algorithm

Download

Upload

Crossed 2 Downloads 2 Uploads

Down Up Down#1 Down#2 Up#1 Up#2

BIC 12.15 6.0 17.0 16.0 10.0 10.0 11.0 10.30

New Reno 7.0 6.40 12.0 7.0 11.0 10.0 10.0 10.0

Vegas 10.0 5.10 13.10 5.20 13.30 14.10 10.0 9.30

Westwood+ 11.30 12.10 47.0 16.0 67.30 68.0 15.10 14.50

H-TCP 8.30 5.45 13.0 6.10 10.0 9.30 10.10 10.30

focusing on the bandwidth percentage they manage

to exploit.

Therefore, it is evident that, regarding the data

transferring scenario, in the crossed data flow

scenario Westwood+ algorithm exhibited the poorest

bandwidth utilization as both flows captured only

approximately 1Mbyte/sec and completed the

transfer on 47 min and 16 min for download and

upload respectively. On the other hand Vegas and

Reno delivered a considerably higher 3Mbyte/sec

aggregated throughout, followed by HTCP

2,5Mbyte/sec and BIC recording 1,2Mbyte/sec.

Valuable information can be extracted observing

that after the completion of each upload, on the

crossed scenario, the corresponding download flow

probes the channel in order to occupy higher

percentage of the available bandwidth. However,

Westwood+ was unable to take advantage of the

absence of the competing flow as it managed to

increase its throughput by only 200Kbyte/sec in the

remaining 31 min whereas H-TCP and Vegas

improved their throughput by 800Kbyte/sec in 7min

and 8min till session termination respectively, Reno

by 600Kbyte/sec in 5min and BIC by 200Kbyte/sec

in 1min.

The clear performance degradation of the

network when using Westwood+ algorithm, as

opposed to the rest of the considered TCP

algorithms when operating on the hybrid topology,

is represented by high required transfer periods and

low measured throughput for all the transport

scenarios. This fact raised some concerns about the

effectiveness of its bandwidth estimation

mechanism. Especially on the crossed scenario

where data and ACKs flow in the same direction, the

result was very high RTT values leading to even

lower bandwidth estimation. Therefore, based on

understanding the theoretical approach of

Westwood+ TCP algorithm and presented

measurements, in the following section

modifications in key parameters and respective

performance evaluation are presented towards

improving Westwood+ performance in the specific

scenarios.

5 PROPOSED MODIFICATIONS

AND EVALUATION

The proposed Westwood+ algorithm modifications

focus on the respective bandwidth estimation

mechanism. Two parameters of the Linux

implementation are going to be modified each one

toward a different goal. The main objectives are,

tackling poor bandwidth estimation due to high

measured RTT values, and slow response of the

algorithm following the absence of the competing

data flow. Both parameters are modified in the

“tcp_westwood.c” code file of the Linux kernel.

The first modification concerns the value of the

constant parameter

TCP_WESTWOOD_RTT_MIN. The

initial value of this parameter is 50msec while the

value considered in experiments presented in this

section is 40msec. The observation triggering this

modification was the comparative analysis of the

average measured value of RTT for Westwood+ as

opposed to the rest algorithms on the hybrid

topology “Fig. 7”. This parameter is used by the

function

westwood_update_window in order to

decide whether the

westwood_filter function that

evaluates the bandwidth samples should be called.

Our goal is to reduce the values of the measured

RTT samples so as the algorithm to achieve better

utilization of the channel’s bandwidth.

The second modification aims to affect the

sensitivity of the Low pass filter (Grieco, 2004),

implemented by the function

westwood_do_filter

using constant coefficients. We are going to present

experiments configured to two different values for

the factor C=7 that is used in the function. The

objective of these changes is to make the algorithm

more aggressive when it comes to probing for

available bandwidth.

PERFORMANCE EVALUATION OF TCP ALGORITHMS ON HYBRID WIRED/WIRELESS LAN TEST-BED

Figure 8: RTT of “1 Upload Scenario” on hybrid topology,

RTTmin modified.

“Fig. 8” shows the RTT measurements for the two

values of the TCP_WESTWOOD_RTT_MIN

parameter. The grey graph represents the initial

50msec value and the black graph the new 40msec

value. As it is depicted, there has been a significant

reduce of 50% in the average RTT value from

100msec to 50msec, which caused a 250% raise on

the used throughput from 1,2Mbyte/sec to

3Mbyte/sec compared to the original Westwood+

algorithm.

Figure 9: Throughput of “Crossed Scenario” on hybrid

topology, factor C modified.

In “Fig. 9” the crossed scenario throughput graphs

for three different values of the constant factor C are

presented. For C=9 the modified Westwood+

algorithm managed to increase its throughput by

400Kbyte/sec in 3,30min while for C=5 by

700Kbyte/sec in 4,10min. Due to the linear pattern

of the graphs it is possible to quantitatively estimate

the algorithm’s aggressiveness increase effect

through the respective measurements’ time duration

decrease down to 1 min. Thus, compared to the

12.5Kbyte/sec/min of the initial C=7 (estimating

that the download flow reached its maximum value

in 16

minute of the experiment) respective results

were increased 9 times for C=9 and 13 times for

C=5. Thus the modified algorithm appears

significantly more aggressive in grabbing the

available bandwidth after the absence of the opposite

data flow and indicating that the default value leads

to a rather passive behaviour.

6 CONCLUSIONS

This paper presents a comparative performance

evaluation of five different congestion control

algorithms using a local area network test-bed setup

comprising by both wired and wireless segments.

The evaluated algorithms have been tested in a

series of experimental, data transferring scenarios

taking considering both wired and hybrid

wired/wireless topologies.

Additionally experiments carried out indicated

possible directions towards improving the

performance of specific TCP algorithms. In this

context it is observed that in hybrid topologies

upload data flows tend to be more aggressive in

capturing the available bandwidth. In the two

parallel flow scenarios TCP manages to evenly

distribute the channel’s available bandwidth to the

two competing flows, but is unable to do so in the

two crossed flow scenarios. On the wired topology,

the two crossed flows converge rapidly to a fair

distribution of the channel’s bandwidth but such a

convergence was not observed in the hybrid

topology.

Additionally, in respective experiments

Westwood+ algorithm recorded the poorest

utilization of the available bandwidth in all data

transfer scenarios assuming wired-wireless topology.

Based on this observation, two modifications are

advocated concerning key parameters of the

implementation code in order to improve the

performance of the bandwidth estimation

mechanism. Respective measurements using the

modified TCP implementation have shown

significant improvement in the algorithm’s

sensitivity as well as bandwidth utilization.

DCNET 2010 - International Conference on Data Communication Networking

REFERENCES

P. Mohaparta, S. Krishnamurthy, "AD HOC NETWORKS

Technologies and Protocols", 2005 Springer Science +

Business Media, Inc. eBook ISBN 0-387-22690-7.

Lisong Xu, Khaled Harfoush, and Injong Rhee, “Binary

Increase Congestion Control for Fast, Long Distance

Networks” In: IEEE Infocom 2004.

S. Mascolo, et. al “TCP Westwood: end-to-end bandwidth

estimation for efficient transport over wired and

wireless networks” In: Proceedings of the ACM

Mobicom 2001, Rome, Italy, July 2001.

Luigi A. Grieco and Saverio Mascolo, “Performance

Evaluation and Comparison of Westwood+, New

Reno, and Vegas TCP Congestion Control”, ACM

SIGCOMM Computer Communication Review, 2004.

V. Jacobson, “Congestion avoidance and control” In:

Proceedings of the ACM Sigcomm’88, Stanford, CA,

August 1988, pp. 314–329.

Floyd, S., Henderson, T. “New Reno Modification to

TCP's Fast Recovery”, RFC 2582, April 1999.

Brakmo, L. S., O’Malley, S.W., and Peterson, L. “TCP

Vegas: End-to end congestion avoidance on a global

Internet”. IEEE Journal on Selected Areas in

Communications, 1995.

D. Leith, R. Shorten, “H-TCP: TCP for high-speed and

long-distance networks”, Hamilton Institute, NUI

Maynooth, 2004.

PERFORMANCE EVALUATION OF TCP ALGORITHMS ON HYBRID WIRED/WIRELESS LAN TEST-BED