AN EXPERIMENTAL STUDY ON THE PERFORMANCE AND

FAIRNESS OF LOSS DIFFERENTIATION FOR TCP

Johan Garcia and Anna Brunstrom

Karlstad University

Universitetsg 2, SE-651 88 Karlstad, Sweden

Keywords:

Loss differentiation, TCP, fairness, multiple connections, performance evaluation

Abstract:

This paper investigates the performance and fairness of receiver-based loss differentiation for TCP. Exper-

iments have been performed with a FreeBSD kernel implementation. As expected, the results verify the

effectiveness of receiver-based loss differentiation when corruption losses are present. However, if the the cor-

rupting link is shared with users that do not employ loss differentiation, the performance gain typically comes

at the expense of increased unfairness. The results show that a single loss differentiating user may in some

cases reduce the bandwidth of users without loss differentiation with up to 35 percent, but there are also cases

where loss differentiation has limited fairness implications. The results further show that if a user employs

multiple TCP connections over the corrupting link the negative effects of corruption losses are reduced. This

is true even if all connections employ regular TCP. Hence, multiple connections at the application level can to

some extent be used as a simple mechanism to limit the impact of corruption losses.

1 INTRODUCTION

Packet switched networks by design provide re-

sources in a dynamic way. In the absence of ad-

mission control, packet switched networks need to

ensure that the trafﬁc does not saturate the network

completely and create a congestion collapse. To con-

trol congestion in the Internet, the Transmission Con-

trol Protocol (TCP) (Postel, 1981) employs conges-

tion control. TCP congestion control interprets packet

loss as congestion and applies some congestion avoid-

ance behavior, typically by the TCP sender halving

the congestion window (Allman et al., 1999).

TCP’s mechanisms for congestion control have

been used with considerable success. There are,

however, circumstances where losses occur that are

not caused by congestion. When congestion avoid-

ance is incorrectly applied also for non-congestion

related losses, TCP performance is degraded. Proto-

cols with similar congestion control, such as the Data-

gram Congestion Control Protocol (DCCP) (Kohler

et al., 2004) and the Stream Control Transfer Proto-

col (SCTP) (Stewart, 2000), share this basic problem.

One important communication environment in which

non-congestion related losses may occur is wireless

networks. Two major causes of non-congestion re-

lated losses in this environment are wireless link er-

rors and mobility induced losses resulting from the

inability to do perfect hand-offs. In this paper we fo-

cus on wireless link errors.

A common approach to handle wireless link errors

is to augment the lower layers with forward error cor-

rection (FEC) and/or link-layer retransmissions. This

is done in many operational wireless networks. As

perceived by TCP, link-layer retransmissions trans-

form the unreliability of the wireless medium into

round-trip time variations. In some instances these

variations can cause spurious timeouts which incur

both an unnecessary retransmission and a reduction

of the congestion window. The D-SACK (Floyd et al.,

2000) and TCP Eifel (Ludwig and Katz, 2000) algo-

rithms provide solutions that can detect the spurious

retransmissions and undo the unnecessary window re-

duction. Although link-layer retransmissions effec-

tively shield TCP from bit errors over the wireless

channel, they are not practical in all situations and

hence there are occasions when TCP is subjected to

non-congestion losses resulting from the traversal of

a wireless link.

A considerable amount of research has been done

on solutions to the performance degradation problem

that occurs when wireless link errors are misinter-

410

Garcia J. and Brunstrom A. (2004).

AN EXPERIMENTAL STUDY ON THE PERFORMANCE AND FAIRNESS OF LOSS DIFFERENTIATION FOR TCP.

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 410-419

DOI: 10.5220/0001403904100419

 SciTePress

preted as congestion by TCP. Examples include us-

ing split connections (Bakre and Badrinath, 1995),

TCP-aware local retransmissions (snoop) (Balakrish-

nan et al., 1995) or loss differentiation at the transport

layer (Balan et al., 2002; Cen et al., 2002; Garcia and

Brunstrom, 2002; Kim et al., 1999) which is the focus

of this paper.

Loss differentiation is the process of discriminat-

ing between different loss causes, making it possible

to differentiate between congestion losses and other

types of losses. Loss differentiation makes it pos-

sible to treat congestion losses and losses caused by

bit errors on a wireless link differently, not applying

the standard congestion control behavior for wireless

link losses. The work presented in this paper exam-

ine different aspects of loss differentiation using an

actual TCP implementation modiﬁed to use loss dif-

ferentiation. The performance gains of using receiver

based loss differentiation are examined. The results

show that the gains are considerable in most cases,

which is consistent with previous studies. Further,

we relate these results to the performance of our sim-

ple multiple connections approach. The results from

these experiments show that the multiple connections

approach considerably improve the throughput when

corruption losses are present. In some cases, multi-

ple connections perform better than loss differentia-

tion. Finally, we examined the fairness aspects of loss

differentiation. The results show that when a corrupt-

ing link is shared between a mix of normal TCP user

ﬂows and loss differentiating user ﬂows, the gain for

the loss differentiation ﬂows in some instances came

at the expense of a noticeable throughput reduction

for the competing normal TCP ﬂows.

The outline of this paper is as follows. In the next

section a background on loss differentiation is pro-

vided followed by a section presenting the results. Fi-

nally, the conclusions of the work are presented.

2 LOSS DIFFERENTIATION

As discussed above, loss differentiation is the pro-

cess of classifying a loss as being caused either by

congestion in the network or by some other event,

typically corruption of some bits by a wireless link.

This section presents a selection of loss differentia-

tion schemes, grouped together according to where in

the network support for loss differentiation is placed:

within the network infrastructure, at the sender-side

or at the receiver-side. The placement typically inﬂu-

ences the loss differentiation precision that can be ob-

tained. The precision of a loss differentiation scheme

describes how well the scheme distinguishes between

congestion losses and wireless losses.

Loss differentiation schemes that rely on infras-

tructural support typically require some changes to

the wireless link base station. Although they may

demand considerable modiﬁcations, the precision of

these schemes is generally very good. Examples

include (Cobb and Agrawal, 1995) who describe a

solution in which base stations send ﬁrst-hop and

last-hop acknowledgments (acks) for the trafﬁc go-

ing from and to a mobile host, respectively. The ex-

tra acks allow the sender to infer where the loss oc-

curred, and also works for mobile-to-mobile commu-

nication. Another example is the syndrome approach

by (Chen et al., 2002) which requires that the base sta-

tion counts the number of packets sent per ﬂow, and

then forwards this number to the receiver in a TCP

option. The receiver can then infer if a packet loss oc-

curred on the wireless link. By employing the support

of routers, the CETEN approach described by (Krish-

nan et al., 2002) notiﬁes the sender of the cumulative

non-congestion losses along a path.

Sender-based loss differentiation schemes require

modiﬁcation only to the sending end-host and are

based solely on processing the incoming ack stream

to infer the network state. Losses occurring when

the network is in a congested state are then classi-

ﬁed as congestion losses. If the network, as per-

ceived by the sender, is in a non-congested state,

losses are classiﬁed as wireless. However, relying on

the ack stream introduces considerable uncertainty.

This uncertainty translates to low loss differentia-

tion precision. The precision also varies with fac-

tors such as trafﬁc load, reordering, amount and dis-

tribution of cross trafﬁc, buffer sizes, link delays,

queue sizes and mechanism parameterization. Ex-

amples include (Kim et al., 1999) who suggest the

use of Linear Increase/Multiplicative Decrease with

History (LIMD/H). In this scheme the effective trans-

mission rate history and congestion throttling his-

tory are maintained to improve precision. Depend-

ing on the relation of the current rate to the effective

rate, losses are classiﬁed into one of three categories:

congestion loss, probe loss, or non-congestion loss.

Another example is (Barman and Matta, 2002) who

present a scheme named NewReno-FF. This scheme

uses an adaptive Flip Flop ﬁlter for the ack round-trip

times (RTTs) in addition to the usual TCP exponential

weighted moving average ﬁlter. The underlying as-

sumption is that for packets that suffer random losses,

the observed RTT varies signiﬁcantly more than when

congestion loss occurs.

Receiver-based loss differentiation schemes require

changes only to the receiver side in order to per-

form loss differentiation. They do not share the sen-

sitivity of the sender-based predictors to the back-

channel conditions. However, since they are receiver-

based, they require a loss notiﬁcation mechanism

to convey the loss differentiation information to the

sender. The precision of receiver based schemes is

AN EXPERIMENTAL STUDY ON THE PERFORMANCE AND FAIRNESS OF LOSS DIFFERENTIATION FOR TCP

411

typically very favorable compared to the sender-based

schemes. Examples include (Cen et al., 2003) who

suggest schemes which are based on the time relations

of incoming packets.They also consider the fairness

aspects of using loss differentiation. Another class

of receiver based schemes are based on the observa-

tion that wireless corruption often do not cause the

whole packet to be lost, but rather a number of bits

in the packet become corrupted. These corrupt bits

will then lead to checksum failure. When a check-

sum control function detects an erroneous checksum

the packet is discarded. The fact that data was dis-

carded due to a checksum error is however not known

outside the checksum control function. By extend-

ing the checksum control the discarded packet can be

classiﬁed as a wireless loss and not congestion. Ex-

amples include (Balan et al., 2002) who describe TCP

HACK, a scheme that uses a TCP option containing

a checksum for the TCP header. When a corrupted

packet arrives, a correct header checksum guarantees

the integrity of the header information which is used

to map the corrupted packet to the correct ﬂow. An-

other example is (Garcia and Brunstrom, 2002) who

discuss a similar approach but without the require-

ment of a checksum option.

Of particular interest is the receiver based loss dif-

ferentiation used in the Datagram Congestion Con-

trol Protocol (DCCP) (Kohler et al., 2004). DCCP

is a transport protocol that provides an unreliable but

congestion-controlled ﬂow of datagrams. DCCP al-

lows for ﬂexibility as to the details of the conges-

tion control by using congestion control identiﬁers

(CCID) that govern the exact behavior of the conges-

tion control. Two different CCIDs are currently being

proposed; TCP-like congestion control (CCID 2) and

TFRC congestion control (CCID 3). DCCP has a ﬂex-

ible checksumming mechanism in the header which

can be set to cover only header information, header in-

formation and partial data or the whole packet. In ad-

dition to the checksum in the header, DCCP also has

a data checksum option. If the data checksum option

is used, a packet having an incorrect data checksum

and a correct header checksum can be classiﬁed as a

corruption loss. For DCCP, corruption errors that do

not affect the header can thus be detected and mapped

to the correct ﬂow. For the TCP-like congestion con-

trol, loss notiﬁcation is then performed by the data

dropped option of DCCP, which allows the receiver to

specify that packets were dropped due to corruption.

Loss differentiation in DCCP with TCP-like conges-

tion control is thus quite similar to the scheme used in

the experiments in this paper as discussed in the next

section.

Although this paper speciﬁcally reports on exper-

iments on the performance and fairness of loss dif-

ferentiation, we note that there are other schemes that

consider the problem of wireless losses for TCP. They

do not explicitly use loss differentiation, but rather

rely on bandwidth estimation techniques to improve

performance. Examples of such work include TCP

Westwood+ (Mascolo et al., 2004), WTCP (Sinha

et al., 2002) and TCP Real (Zhang and Tsaoussidis,

2001). TCP Real has lately been extended with loss

classiﬁcation functionality (Zhang and Tsaoussidis,

2004).

3 EXPERIMENTAL SETUP

This paper reports on performance and fairness exper-

iments performed using a real TCP implementation.

In order to perform experiments a checksum-based

loss differentiation mechanism was implemented in

the FreeBSD 4.5 kernel. In short it works as follows

When a corrupted packet reaches the IP-input func-

tion, the IP header checksum is controlled. If the

checksum is invalid, the packet is discarded since the

IP-addresses are needed to map the packet to a ﬂow.

In this case, the loss is implicitly classiﬁed as a con-

gestion loss. If the IP header checksum is correct, the

packet is forwarded to the TCP input function. Since

it is a corrupted packet the TCP checksum will be er-

roneous, and the packet would normally be discarded.

However, when checksum based loss differentiation

is used, a ﬂow match is attempted before discarding

the packet. A ﬂow match uses the IP-addresses and

the port numbers. After the ﬂow match the sender

is notiﬁed that a corruption loss has occurred via the

use of a TCP-option. The implementation provides

a high precision since it can classify losses accurately

as long as there are no bit errors in the sensitive header

ﬁelds. The exact probability of classifying a corrup-

tion loss as a congestion loss depends on the bit error

distribution and the relation between the header size

and packet size, but in the experiments presented in

this paper the probability is below 0.03. There is no

possibility of falsely classifying a congestion loss as

a corruption loss.

When performing the experiments a setup consist-

ing of a client, a server and a router/emulator was used

as illustrated in Figure 1. The router/emulator ran the

Dummynet software (Rizzo, 1997) to emulate various

link conditions. Three different bandwidths (1Mbps,

384Kbps, 64Kbps), three different delays (10ms, 50

ms, 200 ms), and uniformly distributed bit errors with

seven bit error rates (BER) varying between 0 and

1.2 ∗ 10

−5

were used. The BER values were chosen

so that they correspond to packet loss ratios (PLR) of

approximately 1, 2, 5, 7.5, 10 and 15 percent for the

This paper focuses on performance and fairness aspects

of receiver-based loss differentiation so we only provide a

short description. Further discussion of relevant details are

available in (Garcia and Brunstrom, 2002).

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

412

Control

Script

TCPdump

Ethernet

TCP

TCPdump

Ethernet

TCP

= Control path

= Data path

Router/Emulator

FreeBSD 4.6.2FreeBSD 4.5 FreeBSD 4.5

100Mbps

Client Server

Data sink Data src

100Mbps

P4−2.4GHz P4−2.4GHzP4−2.4GHz

Dummynet

Figure 1: Physical experiment setup

packet size used (1500 bytes). To be noted is that we

used a modiﬁed version of the Dummynet FreeBSD

kernel code that enabled us to control the bit errors

using bit error pattern ﬁles. The same pattern ﬁle was

reused when testing with and without loss differen-

tiation, thus creating identical positioning of the bit

errors for both cases. To avoid the risk of a speciﬁc

bit error pattern skewing the results, 30 different, ran-

domly generated, bit error pattern ﬁles were used for

each parameter combination. The receiver window

was set to its default value of 64kB.

Two test scenarios were set up. In Scenario 1 the

focus was on evaluating the performance improve-

ments for a single ﬂow and conﬁrm and expand on

the positive results reported in earlier studies. The

Scenario 1 setup measured the throughput of a sin-

gle TCP ﬂow. Figure 2 show the logical setup and a

summary of the parameters for scenario 1. One mea-

surement set was performed with a normal TCP ﬂow

and one set with a TCP ﬂow using loss differentia-

tion. The buffer-size in the router was set to 99 pack-

ets, which is larger than the receiver window. Hence

there is no congestion losses in this scenario, and all

losses will be caused by bit errors on the link.

Scenario 2, which is illustrated in Figure 3, was

setup to examine both the throughput of multiple con-

nections and the fairness effects of loss differentia-

tion. To examine the throughput of multiple connec-

tions between the same endpoints ﬁve ﬂows (1Mbps

& 384 Kbps) or two ﬂows (64Kbps) were started si-

multaneously. For these experiments the buffer-size

in the router was set to 99 packets to be comparable

to the single ﬂow case.

When a wireless link is shared between ﬂows, some

of which employ loss differentiation and some which

or TCP Server

TCP−Lossdiff

or TCP Client

One−way delay: 10, 50, 200 ms

Lossrate: 0 − 1.9*10−5 (BER)

Bit error patterns: 30 different w/

uniform distribution

BW: 1 Mbps, 384, 64 Kbps

Buffer: 99 packets

Figure 2: Scenario 1 logical experiment setup

or TCP Server

TCP−Lossdiff

or TCP Client

or TCP Server

TCP−Lossdiff TCP−Lossdiff

or TCP Client

TCP Client

TCP Server

uniform distribution

TCP Client

One−way delay: 10, 50, 200 ms

Lossrate: 0 − 1.9*10−5 (BER)

Bit error patterns: 30 different w/

BW: 1 Mbps, 384 Kbps

For 1 Mbps and 384 Kbps

For 64 Kbps

Buffer: 20, 99 packets

TCP ClientTCP Server

uniform distribution

Bit error patterns: 30 different w/

One−way delay: 10, 50, 200 ms

Lossrate: 0 − 1.9*10−5 (BER)

BW: 64 Kbps

Buffer: 8, 99 packets

Figure 3: Scenario 2 logical experiment setup

do not, there is a possibility that the loss differenti-

ating ﬂow could unfairly decrease the throughput of

the other, non loss differentiating, ﬂows. To exam-

ine this, ﬁve ﬂows (1Mbps & 384 Kbps) or two ﬂows

(64Kbps) were started simultaneously and competed

for the available link capacity. The ﬁrst set of mea-

surements had four or one normal TCP ﬂow(s) com-

peting against a normal TCP ﬂow. The second set had

four or one normal TCP ﬂows competing against a

ﬂow with loss differentiation. The router buffer-size

was set to 20 packets (1Mbps & 384 Kbps) or 8 pack-

ets (64Kbps) thus causing buffer overﬂow, resulting

in congestion losses.

For all experiments 1 Mbyte of data was transferred

to get each throughput value. The throughput values

for the 30 different bit error patterns were for each

combination of parameters averaged to a mean value

and a 95% conﬁdence interval was calculated.

AN EXPERIMENTAL STUDY ON THE PERFORMANCE AND FAIRNESS OF LOSS DIFFERENTIATION FOR TCP

413

4 RESULTS

4.1 Effectiveness of loss

differentiation

By utilizing loss differentiation to distinguish corrup-

tion losses from congestion losses improved perfor-

mance is expected. The results for the Scenario 1 ex-

periments are shown in Figures 4-10. Note that the

relative positions of the data points are comparable

between ﬁgures since the scale is normalized to the

maximum throughput for the different bandwidths.

Due to space restrictions two of the 200ms ﬁgures are

not shown. The mean throughput of one normal TCP

ﬂow is shown by the ﬁlled square marks and the mean

throughput of one TCP ﬂow with loss differentiation

(called ld ﬂow in the ﬁgures) is shown by the ﬁlled

circle marks. The 95% conﬁdence intervals are also

included in the ﬁgures.

As expected, regular TCP pays a considerable per-

formance penalty in the presence of bit-errors. The

positive effect of loss differentiation is evident look-

ing at the higher throughput obtained by the ﬂow em-

ploying loss differentiation. Loss differentiation is

clearly beneﬁciary, and the relative beneﬁt increases

as the loss rate increases. A general trend is that reg-

ular TCP becomes more sensitive to losses at high

bandwidths (BW) and, to a lesser extent, at high link

delays (D). Both increased bandwidth and increased

delay lead to an increase in the bandwidth-delay prod-

uct (BW*D). A larger BW*D requires a larger aver-

age congestion window size to keep the link fully uti-

lized. The requirement of a larger average window

makes the ﬂow more sensitive to losses. The trends

visible in these results conﬁrm the positive results in

previous simulation based studies such as (Krishnan

et al., 2002) and (Cen et al., 2003).

The experimental results discussed above were ob-

tained with a large router buffer (99 packets). Conse-

quently, no congestion losses occurred since the re-

ceiver window restricted the sender before a router

overﬂow could occur. In addition to these experi-

ments, we also performed experiments with smaller

buffer sizes. This lead to the occurrence of both con-

gestion and corruption losses. The results showed that

the buffer size difference in general had little effect,

but the additional losses generated by the buffer over-

ﬂows lowered the throughput somewhat. This was

most visible for the loss differentiating ﬂow. This

ﬂow experienced the most buffer overﬂows since the

normal ﬂow in most cases did not have sufﬁcient

throughput to cause a buffer buildup. Figure 11 shows

the results for the 1 Mbps and 10 ms case with a buffer

size of 20 packets. These results are comparable with

the results for a large buffer in Figure 4. Examination

of the graphs reveal only small differences that can be

attributed to the reduced buffer size and the resulting

additional congestion losses.

4.2 Effectiveness of multiple

connections

As seen in the previous section, there is a large perfor-

mance gain in using loss differentiation. This subsec-

tion focus on an alternate way to address the prob-

lem of decreased throughput when traversing wire-

less links, namely to open multiple connections be-

tween the same end-points. This approach is com-

monly used by web browsers to speed page load times

when downloading multiple objects on a web page.

We propose an adaptation of this approach, to encom-

pass general data transfer, and with the aim of improv-

ing the throughput over a corrupting link rather than

improving page load times. In this presentation we fo-

cus on the performance aspects of this approach. We

do not consider the more practical aspects such as ap-

plication layer re-assembly and other issues related to

modiﬁcation of the applications to use multiple con-

nections instead of single connections. We do, how-

ever, note that this approach can be realized by appli-

cation layer modiﬁcations only, whereas the loss dif-

ferentiation approach discussed in the previous sec-

tion requires changes at the transport and lower lay-

ers.

The results when using multiple connections are

shown as empty squares in Figures 4-10. As evident

in the graphs there is a considerable gain in using mul-

tiple connections, in some cases it is even more ad-

vantageous to use multiple connections than to use a

single loss differentiating ﬂow. When we examine the

reason for this improved performance, we ﬁnd that it

is twofold; additive stream throughputs and additive

initial congestion window.

As stated by (Mathis et al., 1997), the throughput

of a TCP ﬂow is related to the amount of losses. Rel-

evant in this context is that the throughput T CP

is limited by the error rate. If the corruption error

rate is such that the throughput of a ﬂow is limited

by the corruption errors and not the link bandwidth,

then using n multiple connections allows an aggre-

gate throughput closer to the link bandwidth, or up to

n ∗ T CB

. The throughput of each ﬂow thus adds

to the total aggregate throughput that can be achieved

by multiple connections. This improved throughput

is visible in Figures 4-7 and Figure 10. If the single

ﬂow throughput is not severely limited by corruption

errors, then the relative gain is not so large. This is the

case in Figures 8-9 where the single ﬂow bandwidth

is limited by the link capacity and not by corruption

errors.

The other effect improving the results for multiple

connections is the additive initial congestion window.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

414

200

400

600

800

1000

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

5 flows

Figure 4: Scenario 1 results (1Mbps, 10ms)

200

400

600

800

1000

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

5 flows

Figure 5: Scenario 1 results (1Mbps, 50ms)

100

150

200

250

300

350

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

5 flows

Figure 6: Scenario 1 results (384kbps, 10ms)

100

150

200

250

300

350

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

5 flows

Figure 7: Scenario 1 results (384kbps, 50ms)

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

2 flows

Figure 8: Scenario 1 results (64kbps, 10ms)

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

2 flows

Figure 9: Scenario 1 results (64kbps, 50ms)

AN EXPERIMENTAL STUDY ON THE PERFORMANCE AND FAIRNESS OF LOSS DIFFERENTIATION FOR TCP

415

100

150

200

250

300

350

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

5 flows

Figure 10: Scenario 1 results (384kbps,200ms)

For a single ﬂow the initial congestion window is nor-

mally 2 segments, which is then doubled each round

trip time during slow start. Until the window is large

enough to fully utilize the link, (i.e. window>BW*D)

the link is underutilized. For n multiple connections

that are started concurrently, the aggregate initial con-

gestion window is instead n ∗ 2. Consequently, this

window reaches full utilization sooner, leaving the

link underutilized for a shorter period. The inﬂuence

of this effect is dependent of the length of the transfer,

having greater impact on shorter connections. In the

experiments 1Mbyte was transferred, and for some

of the measurements the effect was large enough to

manifest itself clearly in the graphs. For 1Mbps and

384Kbps the throughput of multiple connections is

actually better than for a loss differentiating connec-

tion for small bit error rates. The difference is larger

for the longer delays, which is consistent with the

above discussion. Larger delay gives larger BW*D

which requires a larger congestion window to keep

the link from being underutilized. At startup, a larger

required window takes longer to achieve for a sin-

gle connection than for multiple connections, which

leaves the link underutilized for a longer period for

the single connection.

Since employing multiple connections can be done

at the application level, this approach may in some

cases be preferable compared to modifying the net-

work stack to employ loss differentiation. However,

this approach also makes the end host behavior more

aggressive compared to using a single ﬂow as dis-

cussed in the end of the next section. The per-

formance and fairness of multiple TCP connections

together with non-congestion losses has previously

been examined by(Hacker et al., 2002). However,

they used simulation to examine the effect of non-

congestion losses on large bulk TCP transfers in the

high performance Abilene network, rather than wire-

less networks causing corruption losses.

200

400

600

800

1000

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow

1 ld flow

5 flows

Figure 11: Scenario 1 results with smaller buffer (1 Mbps,

10ms)

4.3 Fairness of loss differentiation

The previous two subsections focused on evaluating

the performance when a single user is traversing a

corrupting link. Another case is when the corrupting

link is shared by several users who compete for the

available link bandwidth. In this case there may be

some users who employ loss differentiation and some

who do not. The fairness effects of using loss differ-

entiation are thus relevant and are highlighted by the

results for the Scenario 2 measurements, as shown in

Figures 12-18. A total of ﬁve ﬂows (or two for the

64Kbps case) were simultaneously competing for the

link bandwidth. The graphs show the ﬁve ﬂows as

groups of four normal TCP ﬂows plus a single ﬂow.

The single ﬂow can be either a normal TCP ﬂow or a

loss differentiating ﬂow. Each graph shows one set of

measurements where the single ﬂow is a normal ﬂow

competing against the four normal ﬂows, or where the

single ﬂow is a loss differentiation ﬂow competing

against the four normal ﬂows.

Using Figure 12 as example, it is visible that

the throughput of a single normal TCP ﬂow (ﬁlled

squares) decrease when the bit error ratio increase.

The four competing normal TCP ﬂows (empty

squares) also decrease their aggregate throughput as

the bit error ratio increase. The set of measurements

using an ld ﬂow shows quite different results. The

throughput of the single ld ﬂow (ﬁlled circles) actu-

ally increase as the bit error ratio increase. To a cer-

tain extent, this increase might be done without caus-

ing additional unfairness since the competing normal

TCP ﬂows often cannot fully utilize a link when the

bit error rate increases. Bandwidth that could not be

used by the competing ﬂow can instead be used by

the ld ﬂow. As was shown in Section 4.1, an ld ﬂow is

practically insensitive to the bit errors that hamper the

throughput for normal TCP ﬂows. However, as shown

by the empty circles, the aggregate throughput of the

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

416

200

400

600

800

1000

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

4 flows vs. 1 normal flow

1 normal flow

4 flows vs. 1 ld flow

1 ld flow

Reduction for other flows when competing

with a loss differentiation flow

Improvement for one flow

when using loss differentiation

Figure 12: Scenario 2 results (1Mbps, 10ms)

200

400

600

800

1000

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

4 flows vs. 1 normal flow

1 normal flow

4 flows vs. 1 ld flow

1 ld flow

Figure 13: Scenario 2 results (1Mbps, 50ms)

100

150

200

250

300

350

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

4 flows vs. 1 normal flow

1 normal flow

4 flows vs. 1 ld flow

1 ld flow

Figure 14: Scenario 2 results (384kbps, 10ms)

100

150

200

250

300

350

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

4 flows vs. 1 normal flow

1 normal flow

4 flows vs. 1 ld flow

1 ld flow

Figure 15: Scenario 2 results (384kbps, 50ms)

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow vs. 1 normal flow

1 normal flow

1 flow vs. 1 ld flow

1 ld flow

Figure 16: Scenario 2 results (64kbps, 10ms)

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

1 flow vs. 1 normal flow

1 normal flow

1 flow1 vs. 1 ld flow

1 ld flow

Figure 17: Scenario 2 results (64kbps, 50ms)

AN EXPERIMENTAL STUDY ON THE PERFORMANCE AND FAIRNESS OF LOSS DIFFERENTIATION FOR TCP

417

100

150

200

250

300

350

2e-06

4e-06

6e-06

8e-06

1e-05

1.2e-05

1.4e-05

Throughput (kbps)

Bit error ratio

4 flows vs. 1 normal flow

1 normal flow

4 flows vs. 1 ld flow

1 ld flow

Figure 18: Scenario 2 results (384kbps, 200ms)

four normal TCP ﬂows is lowered considerably when

they compete with an ld ﬂow. The difference between

the two aggregate throughputs shows the degree of

unfairness that comes as a result of the ld ﬂow’s abil-

ity to work well over a corrupting link and partly tak-

ing bandwidth from the competeing ﬂows. The un-

fairness effect is dependent on the parameter combi-

nation. The aggregate throughput of the competing

ﬂows is reduced by up to 35 percent in Figure 14,

but there is considerably less reduction in other cases.

These results show that fairness indeed can be an issue

when employing loss differentiation, and highlights

the need to also evaluate the fairness implications of

the various schemes proposed. The somewhat unex-

pected results for a bit error ratio of 0 in Figure 17

is caused by loss synchronization effect between the

two ﬂows in the 8 packet buffer. When bit error losses

are also present the synchronization is broken.

The scenario 2 can also be used to illustrate graphs

the unfairness of the multiple connections scheme.

Assume that two users share a corrupting link. One

user employs the multiple connections scheme to in-

crease his throughput over the corrupting link and

uses four normal TCP connections to transfer his data.

The other user does not employ any scheme to im-

prove throughput over the corrupting link. For this

example, the throughput of the two users would be

as illustrated in Figures 12-18. The ﬁrst user, utiliz-

ing multiple connections, would in this case get the

throughput indicated by the empty squares and the

second, normal, user would get the throughput indi-

cated by the ﬁlled squares. Clearly this is far from

fair sharing of the link and it illustrates the unfairness

of the multiple connections approach when it is used

over a shared bottleneck link. The same unfair be-

havior is however to some extent already present in

the Internet, in this case caused by the multiple con-

nections employed by some web browsers to improve

page load times.

5 CONCLUSIONS

This paper reports on experiments performed with a

FreeBSD kernel implementation of a receiver-based

loss differentiation scheme. With regards to the per-

formance of loss differentiation, the results largely

concur earlier work both concerning the problems that

TCP has when used over wireless links with bit errors,

and the effectiveness of receiver-based loss differen-

tiation. Additionally, experiments were performed

using multiple regular TCP connections working as

one aggregate “virtual” connection to improve the

throughput over a link with bit errors. The results

show that employing multiple regular TCP ﬂows im-

prove the throughput considerably, especially for the

lower bit error rates. In several cases, the multiple

connections approach had better throughput than the

loss differentiation approach when run over the same

bit error generating link. However, the multiple con-

nections approach has severe fairness issues if used

over a bottleneck link shared with other users. Nev-

ertheless, multiple connections can be easily achieved

at the application level without requiring any protocol

stack or other changes, so it is an interesting approach

worthy of further study.

The fairness aspect is also relevant in conjunction

with loss differentiation. Fairness issues arise when

several users share a link with bit errors and some

users employ loss differentiation and some do not.

These effects are studied and the results show that

a single loss differentiating user in some cases may

unfairly reduce the bandwidth of the non loss differ-

entiating users with up to 35 percent. However, there

are also cases where loss differentiation has marginal

fairness implications. All in all, the results highlight

the need to consider the fairness issues of the various

loss differentiation schemes proposed in the literature.

Many of the results in this study are applicable also

to other receiver-based loss differentiation schemes.

In particular, the DCCP protocol can employ a TCP-

like congestion control and a loss differentiation func-

tionally very similar to the implementation examined

in this paper. The presented results should thus also

provide a good indication of the behavior of DCCP

loss differentiation.

REFERENCES

Allman, M., Paxson, V., and Stevens, W. (1999). RFC 2581:

TCP congestion control.

Bakre, A. and Badrinath, B. R. (1995). I-TCP: Indirect TCP

for mobile hosts. 15th International Conference on

Distributed Computing Systems.

Balakrishnan, H., Seshan, S., Amir, E., and Katz, R. H.

(1995). Improving TCP/IP performance over wire-

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

418

less networks. In proc. 1st ACM Int’l Conf. on Mobile

Computing and Networking (Mobicom).

Balan, R. K., Lee, B. P., Kumar, K. R. R., Jacob, L., Seah,

W. K. G., and Ananda, A. L. (2002). TCP HACK: a

mechanism to improve performance over lossy links.

Computer Networks, 39(4):347–361.

Barman, D. and Matta, I. (2002). Effectiveness of loss label-

ing in improving TCP performance in wired/wireless

networks. CS Dept Technical report 2002-016, Boston

University.

Cen, S., Cosman, P., and Voelker, G. (2002). End-

to-end differentiation of congestion and wireless

losses. Proc. Multimedia Computing and Networking

(MMCN2002), San Jose, CA, pages 1–15.

Cen, S., Cosman, P. C., and Voelker, G. M. (2003). End-to-

end differentiation of congestion and wireless losses.

IEEE/ACM Transactions on Networking, 11(5):703–

717.

Chen, W.-P., Hsiao, Y.-C., Hou, J. C., Ge, Y., and Fitz,

M. P. (2002). Syndrome: a light-weight approach to

improving TCP performance in mobile wireless net-

works. Wireless Communications and Mobile Com-

puting, (2):37–57.

Cobb, J. A. and Agrawal, P. (1995). Congestion or cor-

ruption? A strategy for efﬁcient wireless TCP ses-

sions. IEEE Symposium on Computers and Commu-

nications, Alexandria, Egypt, pages 262–268.

Floyd, S., Mahdavi, J., Mathis, M., and Podolsky, M.

(2000). RFC 2883: An extension to the selective ac-

knowledgement (SACK) option for TCP.

Garcia, J. and Brunstrom, A. (2002). Checksum-based

loss differentiation. Proceedings 4th IEEE Confer-

ence on Mobile and Wireless Communications Net-

works (MWCN 2002), Stockholm, Sweden.

Hacker, T. J., Noble, B. D., and Athey, B. D. (2002). The ef-

fects of systemic packet loss on aggregate TCP ﬂows.

Proceedings of the 2002 ACM/IEEE conference on

Supercomputing, Baltimore, Maryland, pages 1–15.

Kim, T., Lu, S., and Bharghavan, V. (1999). Improving con-

gestion control performance through loss differentia-

tion. Proceedings International Conference on Com-

puters and Communications Networks (ICCCN99),

Boston, USA.

Kohler, E., Handley, M., Floyd, S., and Padhye, J. (2004).

Datagram congestion control protocol (DCCP). draft-

ietf-dccp-spec-06.txt, Work in progress.

Krishnan, R., Allman, M., Partridge, C., and Sterbenz, J. P.

(2002). Explicit transport error notiﬁcation for error-

prone wireless and satellite networks. BBN Technical

Report No. 8333, BBN Technologies.

Ludwig, R. and Katz, R. H. (2000). The Eifel algorithm:

Making TCP robust against spurious retransmissions.

ACM Computer Communication Review, 30(1):30–

37.

Mascolo, S., Grieco, L. A., Ferorelli, R., Camarda, P., and

Piscitelli, G. (2004). Performance evaluation of West-

wood+ TCP congestion control. Performance Evalu-

ation, 55(1-2):93–111.

Mathis, M., Semke, J., and Mahdavi, J. (1997). The macro-

scopic behavior of the TCP congestion avoidance al-

gorithm. Computer Communications Review, 27(3).

Postel, J. (1981). RFC 793: Transmission control protocol.

Rizzo, L. (1997). Dummynet: a simple approach to the

evaluation of network protocols. ACM Computer

Communication Review, 27(1):31–41.

Sinha, P., Nandagopal, T., Venkitaraman, N., Sivakumar, R.,

and Bharghavan, V. (2002). WTCP: A reliable trans-

port protocol for wireless wide-area networks. Wire-

less Networks, 8(2):301–316.

Stewart, R. (2000). RFC 2960: Stream Control Transmis-

sion Protocol.

Zhang, C. and Tsaoussidis, V. (2001). TCP Real: Improv-

ing real-time capabilities of TCP over heterogeneous

networks. Proceedings of the 11th IEEE/ACM NOSS-

DAV.

Zhang, C. and Tsaoussidis, V. (2004). Error differentiation

with measurements based on wave patterns. Computer

Communications, 27(10):989–1000.

AN EXPERIMENTAL STUDY ON THE PERFORMANCE AND FAIRNESS OF LOSS DIFFERENTIATION FOR TCP

419