USING ECN MARKS TO IMPROVE TCP PERFORMANCE OVER

LOSSY LINKS

Haowei Bai

Honeywell Labs

3660 Technology Drive, Minneapolis, MN 55418, USA

Mohammed Atiquzzaman

School of Computer Science

University of Oklahoma, Norman, OK 73019-6151, USA

David Lilja

Department of Electrical and Computer Engineering

University of Minnesota, 200 Union St. SE, Minneapolis, MN 55455, USA

Keywords:

Wireless network, explicit congestion notiﬁcation, TCP/IP, congestion control

Abstract:

TCP was designed for wireline networks, where loss events are mostly caused by network congestion. The

congestion control mechanism of current TCP uses loss events as the indicator of congestion, and reduces

its congestion window size. However, when a lossy link is involved in a TCP connection, non-congestion

random losses should also be considered. The congestion window size should not be decreased if a loss event

is caused by link corruptions. To improve TCP performance over lossy links, in this paper, we ﬁrst present that

zero congestion loss could be achieved by appropriately setting the ECN marking threshold in the RED buffer.

Secondly, we propose a new TCP algorithm, called Differentiation Capable TCP (Diff-C-TCP). Diff-C-TCP

makes an assumption that packet losses are caused by link corruptions, and uses ECN (Explicit Congestion

Notiﬁcation) to determine any loss that may occasionally happen due to network congestion. We have shown

that Diff-C-TCP performs very well in the presence of a lossy link.

1 INTRODUCTION

The TCP/IP (Transmission Control Protocol/Internet

Protocol) protocol was originally designed for wire-

line networks. Wireless links have a number of fun-

damentally different characteristics from wired links.

Wireless links are dominated by low bandwidth and

high error rates, which degrade the performance of

TCP (Bai and Atiquzzaman, 2003). In the current

TCP, the congestion control mechanism uses packet

loss as the indicator of congestion, and reduces its

congestion window size. However, when the packet

loss is caused by link errors rather than congestion,

TCP’s reduction of the congestion window size is un-

necessary and degrades its throughput.

In most cases, packet losses due to corruption are

more signiﬁcant than congestion losses when a lossy

link is involved in a TCP connection. In such a case,

TCP may not be able to transmit or receive at the full

available bandwidth, because the TCP algorithm is

unnecessarily wasting time in slow-start or congestion

avoidance procedures triggered by link errors. Con-

sequently, the current congestion control algorithms

in TCP result in very poor performance over lossy

links. Signiﬁcant performance improvements can be

achieved (Samaraweera and Fairhurst, 1998) if losses

due to network congestion and corruption in lossy

wireless links could be appropriately differentiated.

Researchers have applied heuristics, such as loss

predictors, to distinguish between congestion and er-

rors due to link errors (Biaz and Vaidya, 1997; Biaz

and Vaidya, 1998b), but their simulation results indi-

cated that their loss predictors did not perform well.

They proposed a modiﬁed TCP-Reno scheme called

TCP-Aware (Biaz and Vaidya, 1998a) to differentiate

between packet losses due to congestion and losses

due to link errors. However, their scheme works well

only when the last hop for the connection is wireless,

the bandwidth of the wireless link is much smaller

than the bandwidth of the wired link, and the overall

packet loss rate is small (Biaz and Vaidya, 1998a), all

of which are too limited in the real world Internet.

437

Bai H., Atiquzzaman M. and Lilja D. (2004).

USING ECN MARKS TO IMPROVE TCP PERFORMANCE OVER LOSSY LINKS.

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 437-445

DOI: 10.5220/0001405604370445

 SciTePress

Without any other additional information, the ex-

isting implicit loss feedback mechanisms in TCP does

not allow distinguishing between congestion and cor-

ruption losses (Dawkins et al., 2000). ECN (Explicit

Congestion Notiﬁcation) (Floyd, 1994) was proposed

as an explicit indicator of congestion. It can be used

to quickly and unambiguously inform sources of net-

work congestion, without the sources having to wait

for either a retransmit timer timeout or three duplicate

ACKs (Acknowledgements) to infer a lost packet.

For bulk-data connections, this mechanism can avoid

unnecessary packet drops for low-bandwidth delay-

sensitive TCP connections, and can avoid some un-

necessary retransmit timeouts in TCP (Ramakrishnan

and Floyd, 1999). Since ECN is an explicit indica-

tor of network congestion, it provides the possibility

to differentiate two types of losses. If the buffer in a

router is optimally dimensioned and the RED (Ran-

dom Early Detection) threshold in the router buffer

is appropriately set, zero congestion loss could be

achieved by appropriately adjusting the source’s con-

gestion window size based on feedback from ECN

signals. Previous researchers (Kunniyur and Srikant,

2000; Liu and Jain, 2001a; Abouzeid and Roy, 2000)

have done some initial work in this direction as de-

scribed below.

Authors in (Kunniyur and Srikant, 2000) presented

a framework for designing end-to-end congestion

control schemes in a network where each user may

have a different utility function. They considered

ECN marks as an alternative to losses for conges-

tion notiﬁcation. Using this model, they showed that

the ECN marking level can be designed to nearly

eliminate congestion losses in the network by choos-

ing the marking level independently for each node in

the network. However, the drawback of their work

is that they achieved zero congestion loss by over-

provisioning the network.

Authors in (Liu and Jain, 2001a) analyzed the

queue dynamics at the congested router, and derived

the closed-form formula and buffer requirements to

achieve zero loss and full link utilization. However, as

they stated in their work, they did not get the mathe-

matical expression for the average share of bottleneck

link bandwidth which is the most important parame-

ter of their model. Instead, they used simulation to il-

lustrate the relationship between the average share of

bottleneck link bandwidth and the Round Trip Time

(RTT). The same difﬁculty was also encountered by

authors in (Abouzeid and Roy, 2000). As a solution,

they introduced an unknown constant into the ﬁnal ex-

pression.

Motivated by (Liu and Jain, 2001a; Abouzeid and

Roy, 2000), we derive the exact mathematical model

for the average share of bottleneck link bandwidth

by modeling the ECN marking dynamics as a Pois-

son Process. We ﬁnally end up with a comprehen-

sive mathematical model for achieving zero conges-

tion loss. The objective of developing the exact math-

ematical model was that, if we can eliminate all net-

work congestion losses in a heterogeneous network

environment involving lossy links, or if the conges-

tion losses are a small fraction of losses due to link

error, with negligible error all losses can be attributed

to random losses due to link errors. This observation

leads to our proposed Diff-C-TCP which is discussed

in detail in Section 3.2.

Diff-C-TCP assumes loss events indicate link cor-

ruption and uses ECN as congestion indication with

the precondition of zero congestion loss to differenti-

ate between congestion and corruption. As mentioned

earlier, because of link errors, packet losses due to

corruption is more signiﬁcant in a lossy network. To

have a high TCP throughput when a TCP connection

traverses a lossy link, the TCP source should persist

in the previous utilization of bandwidth instead of re-

ducing the transmission rate when the loss is due to

corruption.

The contributions of this paper are as follows:

• We develop a comprehensive mathematical model

to calculate the value of the average share of bot-

tleneck link bandwidth. The model ensures zero

congestion loss in the network.

• Based on the possibility of zero congestion loss, we

Propose and evaluate a new TCP algorithm called

Diff-C-TCP to improve the performance of TCP

over lossy links.

The rest of this paper is organized as follows.

In Section 2, we present a comprehensive model to

achieve zero congestion loss with multiple compet-

ing TCP ﬂows; the model is the basis of our pro-

posed Diff-C-TCP algorithm discussed in Section 3.

In Section 4, we describe the simulation methodol-

ogy that has been used to evaluate the performance of

our proposed algorithm. Performance improvements

achieved by our proposed algorithm as compared to

current TCP are presented in Section 5. Concluding

remarks are ﬁnally given in Section 6.

2 ELIMINATING CONGESTION

LOSSES WITH ECN

If the buffer in the router is optimally dimensioned

RED threshold is appropriately set, zero-loss conges-

tion control could be achieved by appropriately ad-

justing the source’s congestion window size based on

the notiﬁcation by ECN.

In this section, we ﬁrst describe the ECN mecha-

nism and provide the requirements for zero conges-

tion loss in two cases (Liu and Jain, 2001a), viz, one

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

438

TCP connection and multiple competing TCP con-

nections, in Sections 2.2 and 2.3 respectively. The dif-

ﬁculty in using the zero congestion model is the cal-

culation of the average share of bottleneck link band-

width. In this section, we have solved that problem

by deriving the exact mathematical expression for the

average share of bottleneck link bandwidth, the most

important parameter in zero congestion loss model,

by modeling the ECN marking dynamics as a Pois-

son Process in Section 2.4. We ﬁnally end up with a

comprehensive mathematical model for achieving the

zero-loss TCP congestion control.

2.1 Analysis Assumptions

In order to set up a proper but not complicated model

and analyze the queue dynamics, as well as zero-loss

requirements, we make the following assumptions (as

used by authors in (Liu and Jain, 2001a; Liu and Jain,

2001b) for a similar model):

• Senders always have data to send and will send as

many as their windows allow.

• Receiver windows are large enough.

• There are no delayed acknowledgements.

• All packets have the same length.

2.2 Queue Dynamics Analysis with

One TCP Flow

Along the path with only one TCP connection, if the

bottleneck link bandwidth is µ packet/second, then

the downstream packet inter-arrival time and the ac-

knowledgement inter-arrival time on the reserve link

must be greater than or equal to

(Liu and Jain,

2001a).

Based on this statement, we use the analytical

model shown in Figure 1 and notations as follows:

w(t) = The sender’s window size at time t.

r = Round Trip Time (RTT).

µ = Bandwidth of the bottleneck link.

= The time a packet needs to traverse from the

sender to the router.

T = The threshold of RED router.

T = The optimum threshold of RED router for the

purpose of zero congestion loss.

P = The packet which increases the queue length

over T .

From (Liu and Jain, 2001a), we know that if at time

t the bottleneck link has been busy for at least r sec-

onds, and a packet just arrives at the congested router,

the queue length at the congested router is

Q(t) = w(t − t

) − rµ. (1)

r-t

Figure 1: Analytical model of one TCP connection.

Based on this theorem, requirements for achieving

zero congestion loss could be described by the rela-

tionship between the queue length and the threshold

of RED router. Results reported by authors in (Liu

and Jain, 2001a) for the two different phases, slow

start and congestion avoidance, are given below:

2.2.1 Slow Start Phase

The maximum queue length in slow start phase is

2T + rµ + 1.

2.2.2 Congestion Avoidance Phase

The maximum queue length in the congestion avoid-

ance phase is T + 1; the minimum queue length is

T −rµ+1

. To derive the optimal threshold for the RED

queue, the bottleneck link should be fully utilized,

i.e.,

T − rµ + 1

≥ 0, (2)

which means

T ≥ rµ − 1. (3)

However, if T > rµ − 1 indicating that the minimum

queue length is a positive number, the RED router

may have long queueing delays. Thus, the optimal

value of threshold (

T ) for achieving zero congestion

loss, full link utilization and minimum queueing de-

lay in congestion avoidance phase can be calculated

from

T − rµ + 1

= 0. (4)

In other words, the optimal value of threshold is

T = rµ − 1. (5)

Finally, we end up with various cases of possible

threshold as shown in Table 1.

2.3 Queue Dynamics Analysis with

Multiple competing TCP ﬂows

Figure 2 shows the system model considered in this

section. We use the following notations in the analy-

sis:

USING ECN MARKS TO IMPROVE TCP PERFORMANCE OVER LOSSY LINKS

439

Table 1: Value of threshold T for the case of one TCP ﬂow.

T Link Status

T > rµ − 1 The link is fully utilized, but packets suf-

fer larger queueing delay.

T = rµ − 1 Optimal threshold; zero congestion

loss, full link utilization and minimum

queueing delay.

T < rµ − 1 The link is under-utilized.

(t) = The i

sender’s window size at time t.

= Round Trip Time (RTT).

= The average share of the bottleneck link band-

width.

= The time a packet needs to traverse from the

sender to the router.

T = The threshold of RED router.

Router D

Figure 2: Analytical model of multiple competing TCP con-

nections.

The analysis of the case of multiple competing TCP

connections is based on the discussion in Section 2.2.

For the analytical model shown in Figure 2, the queue

length expression given by (Liu and Jain, 2001a) is

shown below.

Q(t) =

i=1

(t − t

) − r

). (6)

As we mentioned in Section 1, the difﬁculty in using

Equation (6) is in calculating µ

. We show the calcu-

lation of µ

in Section 2.4.

Similar to the discussion in Section 2.2, we give the

requirements for achieving zero congestion loss for

two different cases, slow start phase and congestion

avoidance phase. These results are also reported by

authors in (Liu and Jain, 2001a).

2.3.1 Slow Start Phase

It is almost unlikely for multiple TCP connections to

send at the same time. If some connections are in slow

Table 2: The value of threshold T for the case of multiple

TCP ﬂows.

T Description

T > rµ − 1 The link is fully utilized, but packets suf-

fer larger queueing delay.

µ − m ≤

T ≤ rµ − 1

Optimal threshold; zero congestion loss,

full link utiliztion and minimum queueing

delay.

T < rµ − m The link is under-utilized.

start phase, while others are in congestion avoidance

phase, the queue length cannot be increased as fast as

all connections are in slow start phase. If all connec-

tions start simultaneously, this could be considered

as one aggregate ﬂow. Accordingly, the maximum

queue length is still 2T + r

+ 1.

2.3.2 Congestion Avoidance Phase

The maximum queue length in congestion avoidance

phase is T +α, α ∈ [1, m]; the minimum queue length

T −r

+α

, α ∈ [1, m]. Three different threshold

values for the case of multiple TCP ﬂows are shown

in Table 2.

2.4 Calculating the Average Share of

Bottleneck Link Bandwidth with

Multiple TCP Flows

In Section 2.2 and Section 2.3, we have summa-

rized the results of RED threshold and buffer size for

achieving zero congestion loss at a RED gateway as

described in (Liu and Jain, 2001a). However, they

did not obtain the quantitative value of the average

share of bottleneck link bandwidth (¯µ

in Equation

(6) which is the most important parameter in the zero

congestion loss model.

In this section, we show the calculation of the aver-

age share of the bottleneck link bandwidth. We con-

sider the system model shown in Figure 2 and the

Reno version of TCP. To keep the consistence, we use

all the assumptions we made for the previous model

in Section 2.2 and Section 2.3.

Figure 3 shows the window evolution approxima-

tion for TCP-Reno sessions with different round trip

delays sharing a bottleneck link with a RED gateway.

The loss events are represented by ’×’ marks. This

approximation has been used by many researchers

(Abouzeid and Roy, 2000; Abouzeid et al., 2000) for

the analytic understanding of the RED performance.

Based on Figure 3, we use the following notations in

our model.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

440

Window

size

Time

W1,1

W1,2

W2,1

W3,1

W2,2

Figure 3: The window evolution approximation with two

TCP-Reno ﬂows.

i,j

(t) = The j

TCP session’s window size right

before the previous loss event.

i+1,j

(t) = The j

TCP session’s window size

right before the current loss event.

avg

(t) = The time-average window size for the

TCP session.

i,j

= Round Trip Time (RTT) of the j

TCP ses-

sion.

= Average share of the bottleneck link band-

width.

= The time when i

congestion loss event hap-

pens.

Consider the scenario in Figure 3; X

= L

− L

i−1

denotes the inter-loss duration. The window evolution

could be expressed by the following equation.

i+1,j

(t) =

i,j

(t)

i,j

. (7)

Since loss events are determined by both the trafﬁc

type and the random marking at RED router, it is rea-

sonable to consider {X

} as an Independent Identical

Distributed (i.i.d.) renewal process. If we assume in

any length of time interval, the number of loss event

is Poisson distributed, then the total number of loss

events in the interval (0, t) is a Poisson process, de-

noted by N(t). Therefore, the loss time interval X

an i.i.d. exponential random variable with the param-

eter λ. Its probability density function (pdf) is

(t) = λe

−λt

u(t). (8)

In addition, the waiting time T [n] = Σ

k=1

for a

loss event is a gamma distributed random variable

with parameters (n, λ). Its pdf can be found as

(t) =

λe

−λt

(λt)

n−1

Γ(n)

u(t), (9)

which is, in the other form,

(t) =

−λt

n−1

(n − 1)!

u(t). (10)

Based on the mathematical nature of the window

evolution we analyzed above, we ﬁnally calculate the

average share of bottleneck link bandwidth, which has

been deﬁned as

avg

(t)

i,j

. (11)

Taking the expectation for both sides of Equation (7),

we have

i+1,j

(t) =

i,j

(t)

i,j

. (12)

Since any two loss events have the same statisti-

cal characteristics, it is apparent that w

i+1,j

(t) and

i,j

(t) have the same expected value. Thus,

(t) = 2

i,j

. (13)

Recall that the loss time duration is a renewal process

and the total number of loss events during any length

of time interval is a Poisson process, from Equation

(8), we should have

= E[X

] =

. (14)

Therefore,

(t) =

λr

i,j

. (15)

Because Poisson process is ergodic in mean (See

Appendix A for the proof), using the property of er-

godicity, we have

avg

(t) =

λr

i,j

. (16)

Finally, the average share of the bottleneck link band-

width is

avg

(t)

i,j

λr

i,j

. (17)

Remarkably, the above result implies that the con-

nection with the shortest RTT has the largest average

share, which is also found by authors in (Liu and Jain,

2001a; Mistra et al., 1999).

3 USING ECN TO IMPROVE THE

TCP PERFORMANCE OVER

LOSSY LINKS: DIFF-C-TCP

From the previous discussion, we can eliminate all

network congestion losses (or the congestion losses

are a small fraction of random losses) in a hetero-

geneous network environment involving lossy links.

Therefore, with the negligible error all losses can be

attributed to random losses. This leads to our pro-

posed Diff-C-TCP discussed in this section.

USING ECN MARKS TO IMPROVE TCP PERFORMANCE OVER LOSSY LINKS

441

3.1 Design Assumptions

Before we start to illustrate the principle of our pro-

posed Diff-C-TCP algorithm, we make the following

assumptions:

• Our proposed algorithm is used within a WAN or

an enterprise network, where it is possible to make

all routers and end-systems ECN-capable.

• When we mention wireless links, in order to keep

our discussion focused, we do not consider mobil-

ity issues such as handoff or power requirements.

3.2 The Proposed Diff-C-TCP

Our proposed algorithm assumes that packet losses

indicate corruption, and the TCP sender uses ECN as

an explicit notiﬁcation of network congestion. Diff-

C-TCP’s response to ECN is similar to TCP’s re-

sponse to packet losses. In other words, the receipt

of ECN packets should trigger a response to network

congestion. Packet losses are treated as link errors

unless ECN packets are received.

Figure 4 shows the kernel of our proposed Diff-C-

TCP algorithm at the sender’s side. A Diff-C-TCP

sender treats the situation that the retransmit timer

times out without receiving any ECN

ECHO packet

and (or) receiving duplicate acknowledgements as the

indication of link errors. Most often, this is the case

in a network with wireless links (packet losses due

to link errors). In this case, the Diff-C-TCP source

does not decrease cwnd. If the Diff-C-TCP sender re-

ceives the ECN ECHO packet sent by the receiver, the

sender treats it as network congestion and triggers the

Fast Recovery algorithm (Ramakrishnan and Floyd,

1999) as in the current TCP.

In Diff-C-TCP, the congestion window size is ap-

propriately controlled in the presence of either net-

work congestion or corruption. Congestion win-

dow is halved using Fast Recovery algorithm when

there is network congestion (explicitly notiﬁed by

ECN ECHO packets), and persists at the previous

value in the presence of corruption. There are two

mechanisms that might be applied to adjust conges-

tion window when Diff-C-TCP sender detects corrup-

tion: (i) keep cwnd unchanged as the previous value;

(ii) use Congestion Avoidance algorithm to slowly in-

crease cwnd. In our algorithm, we adapt the ﬁrst

mechanism: make congestion window persist in the

previous value.

ECN mechanism will be most effective if it is used

with active queue management (such as RED) (Ra-

makrishnan and Floyd, 1999) as illustrated in Fig-

ure 5. In active queue management, when a buffer

reaches a certain threshold, the router will send a CE

packet to the TCP receiver. Routers send CE packets

before their buffers overﬂow. Therefore, packet drops

ECN_ECHO

Packet?

Time out/DUPACKs

(Assume corruption

happened)

DO NOT decrease the

congestion window size;

Retransmit loss packet

Fast Recovery

(W<=W/2)

Figure 4: States transition diagram of Diff-C-TCP kernel.

due to congestion happen only after the router has sent

CE packets. Upon receiving the CE packet, the TCP

Packet drops

due to buffer

overflows

CWR Packet

ECN_ECHO

Packets

CE Packet

Router

ECN_ECHO

Packets

Active Queue

Management

Threshold

Figure 5: A simple Diff-C-TCP based model.

receiver will keep sending ECN ECHO packet back

to the sender until it receives a CWR packet from the

sender, which means the sender has responded to net-

work congestion. The sender only responds to the ﬁrst

ECN ECHO packet and ignores others up to one RTT.

3.3 Illustration of Diff-C-TCP

Depending on the threshold of RED and the level of

network congestion, ECN ECHO packets can arrive

at the sender either before or after the retransmit timer

times out due to congestion packet losses (caused by

buffer overﬂow). Our proposed Diff-C-TCP is effec-

tive in both the above cases as described below.

3.3.1 Case 1: Retransmit timer times out after

ECN ECHO packets are received by the

sender.

In this case, the sender will respond to congestion as

indicated by the receipt of ECN ECHO packets. This

case is desirable.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

442

3.3.2 Case 2: Retransmit timer times out before

ECN ECHO packets are received by the

sender.

In this case, as shown in Figure 6, the retransmit timer

timeout happens at time t

, and ECN ECHO pack-

ets are received by the sender at time t

. If the dif-

ference between t

and t

is small enough, though

the TCP sender does not respond to packet losses in-

dicated by retransmit timer timeout at t

(according

to our Diff-C-TCP), ECN ECHO packets will arrive

very quickly, which will trigger Fast Recovery mech-

anism to relieve the network out of congestion.

The difference between t

and t

can be decreased

by decreasing t

as described below. As mentioned

above, using active queue management such as RED,

when buffer reaches threshold, it will send the CE

packet to receiver. Upon receiving the CE packet,

the receiver starts to send ECN ECHO packets to the

sender. Though it is difﬁcult to control the travel

time of ECN ECHO packets from the receiver to the

sender, we can make the receiver send ECN

ECHO

packets earlier by letting the router send the CE

packet earlier. The earlier ECN

ECHO packets are

sent, the earlier they arrive at the sender, i.e., the

smaller the value of t

is. The time when the router

sends the CE packet is decided by the value of thresh-

old. Therefore, an optimum value of RED’s threshold

is very important for the sender to receive congestion

notiﬁcation quickly. Optimal RED threshold is one of

our current research topics.

time

t1 t2

t2-t1

Timeout

ECN_ECHO

Figure 6: Time sequence of Case 2.

4 SIMULATION

METHODOLOGY

We have evaluated the performance of our Diff-C-

TCP algorithm using the ns (ns Version 2.1b6) sim-

ulation tool from Berkeley (Berkeley/LBNL, 2000).

The ECN implementation is based on RFC 2481 (Ra-

makrishnan and Floyd, 1999). Our network topology

for conducting simulations is shown in Figure 7.

Two local area network (10 Mbps) are connected

using a lossy wireless link (64Kbps) with a propaga-

tion delay of 280ms. Like previous researchers (Bal-

akrishnan et al., 1997; Parsa and Garcia-Luna-

S D

Router A

Router B

10Mb, 2ms

64Kb, 280ms

10Mb, 4ms

LAN

Figure 7: LAN interconnection using a lossy wireless link.

Aceves, 1999), a Uniform random error model is used

to generate random errors on the wireless link. In-

stead of dropping packets at routers, RED routers are

used in our simulations to set the CE bit in the packet

header.

All the links in Figure 7 are labelled with a (band-

width, propagation delay) pair. The full-duplex link

between router A and router B has a BER (bit-error

rate), which varies between 1e

−7

to 1.2e

−4

in our

simulation. The receiver’s advertised window size,

which is also equal to the initial ssthresh at the sender,

is set to 30 segments. The packet size is set to 1000

bytes (when BER is below 5e

−5

) or 512 bytes (when

BER exceeds 5e

−5

). Ftp was used in our simulation

to transfer data from the source to destination.

5 SIMULATION RESULTS

In this section, we present and compare the goodput

(bit/s) (the amount of useful information being re-

ceived by the receiver per second, excluding errors)

and the normalized throughput of our proposed Diff-

C-TCP with the traditional TCP.

Figure 8 compares the goodput in bit/s of both Diff-

C-TCP and the current TCP with ECN capability. The

normalized throughput of both the TCPs are shown

in Figure 9. We see that Diff-C-TCP’s throughput is

much higher than that of the current TCP with ECN

capability. At a BER of 5e

−5

, the goodput of our Diff-

C-TCP is almost 5 times higher than that of the cur-

rent TCP. From Figures 8 and 9, this improvement

is much higher at higher values of BER. In addition,

the throughput of the current TCP with ECN suffers

more severely than our Diff-C-TCP as the error rate

increases. We can also see that, with the increase of

the value of BER, the throughput of the current TCP

with ECN capability decreases much faster than the

throughput of our Diff-C-TCP. For example, accord-

ing to Figure 8, when the value of BER increases from

−5

to 5e

−5

, the current TCP’s goodput decreases

by 77 % in contrast to our Diff-C-TCP whose good-

put only decreases by 12 %. This is also valid for

normalized throughput as shown in Figure 9.

As mentioned previously, the current TCP with

ECN makes an assumption that every loss event is

caused by network congestion, and a congestion con-

USING ECN MARKS TO IMPROVE TCP PERFORMANCE OVER LOSSY LINKS

443

0 0.2 0.4 0.6 0.8 1 1.2

x 10

−4

x 10

BER

Goodput (bit/s)

Tahoe TCP with ECN capability

Diff−C−TCP

Figure 8: Comparison of goodput (bit/s).

0 0.2 0.4 0.6 0.8 1 1.2

x 10

−4

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

BER

Normalized Throughput

Tahoe TCP with ECN capability

Diff−C−TCP

Figure 9: Comparison of normalized throughput.

trol algorithm is triggered. As a result, the congestion

window size must be reduced. Therefore, each loss

event, regardless of whether it is due to congestion

or corruption, degrades the throughput. With Diff-

C-TCP, all packet losses are assumed to be caused by

link errors, and network congestion is indicated by the

receipt of ECN ECHO packets. The congestion win-

dow will not be changed in the presence of corruption

losses. Thus, only network congestion can affect its

throughput. Furthermore, for the current TCP with

ECN, a higher value of BER results in more packet

drops and more frequent reduction of the congestion

window. Because of this, higher values of BER re-

sult in high frequency of reduction of the congestion

window size. It is therefore almost impossible for the

congestion window size to reach a high value, even

during a non-congestion-loss period.

6 CONCLUSION

We have derived a comprehensive model for achiev-

ing zero congestion loss. Based on the possibility of

zero-loss congestion control, we proposed a new TCP

algorithm (named Diff-C-TCP) to improve the TCP

throughput in the presence of non-congestion related

losses in a lossy wireless environment. Diff-C-TCP

improves the network throughput by assuming that

all loss events are caused by link errors, unless the

network explicitly informs the sources of congestion.

Network congestion is explicitly indicated by the re-

ceipt of ECN ECHO packets.

The proposed Diff-C-TCP has been studied in de-

tail using simulation and has been found to signif-

icantly improve TCP performance in lossy wireless

links. Our simulation results have shown that we have

achieved up to 5 times throughput improvement, over

the current TCP with ECN for data transfer across a

typical wireless link with high bit-error rates.

A THE PROOF OF THE

ERGODICITY OF POISSON

PROCESS

Poisson process can be expressed as (See Figure 10)

N(t)

Figure 10: Poisson process.

N(t) =

∞

u(t − T [n]). (18)

As mentioned in Section 2.4, T [n] =

is the

arrival time for a loss event. Because N(t) is a Pois-

son process, during any time internal length of t, say,

(0, t), the total number of events is Poisson distributed

with mean λt.

Deﬁne a random sequence Y [n] = the total number

of events happened in the time interval (0, t), accord-

ing to the above analysis,

E[Y [n]] = λt. (19)

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

444

Then, according to Strong Law of Large Numbers,

lim

n→∞

k=1

Y [k] = λt. (20)

Thus,

lim

n→∞

k=1

Y [k] = λt = E[N (t)]. (21)

In other words, the time average is equal to the ex-

pected value. Therefore, Poisson process is ergodic

in mean.

REFERENCES

Abouzeid, A. and Roy, S. (2000). Analytic understanding

of RED gateways with mutiple competing TCP ﬂows.

In GLOBECOM, pages 555–560, San Francisco, CA.

Abouzeid, A., Roy, S., and Azizoglu, M. (2000). Stochastic

modeling of tcp over lossy links. In INFOCOM, Tel

Aviv, Israel.

Bai, H. and Atiquzzaman, M. (2003). Error modeling

schemes for fading channels in wireless communica-

tions: A survey. IEEE Communications Surveys and

Tutorials, 5(2):2–9.

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

R. (1997). A comparison of mechanisms for improv-

ing TCP performance over wireless links. IEEE/ACM

Transactions on Networking, 6(5):756–769.

Berkeley/LBNL, V. P. U. (2000). ns v2.1b6: Network sim-

ulator. http://www-mash.cs.berkeley.edu/ns/.

Biaz, S. and Vaidya, N. (1997). Using end-to-end statis-

tics to distinguish congestion and corruption losses: A

negative result. Technical report 97-009, Texas A&M

University.

Biaz, S. and Vaidya, N. (1998a). Discriminating congestion

losses from wireless losses using inter-arrival times at

the receiver. Technical report 98-014, Texas A&M

University.

Biaz, S. and Vaidya, N. (1998b). Sender-based heuristics for

distinguishing congestion losses from wireless trans-

mission losses. Technical report 98-013, Texas A&M

University.

Dawkins, S., Montenegro, G., Kojo, M., Magret, V., and

Vaidya, N. (2000). End-to-end performance implica-

tions of links with errors. draft-ietf-pilc-error-03.txt.

Floyd, S. (1994). TCP and explicit congestion notiﬁcation.

ACM Computer Communication Review, 24(5):10–

23.

Kunniyur, S. and Srikant, R. (2000). End-to-end congestion

control schemes: Utility functions, random losses and

ECN marks. In INFOCOM, pages 1323–1332, Tel

Aviv, Israel.

Liu, C. and Jain, R. (2001a). Improving explicit conges-

tion notiﬁcation with mark-front strategy. Computer

Networks, 35(2-3):185–201.

Liu, C. and Jain, R. (2001b). Improving explicit congestion

notiﬁcation with the mark-front strategy. Computer

Networks, 35(2-3):185–201.

Mistra, A., Ott, T., and Baras, J. (1999). The window dis-

tribution of multiple TCPs with random loss queues.

In GLOBECOM, pages 1714–1726, Rio de Janeiro,

Brazil.

Parsa, C. and Garcia-Luna-Aceves, J. (1999). TULIP: A

link-level protocol for improving TCP over wireless

links. In WCNC, pages 1253–1257, New Orleans,

Louisana.

Ramakrishnan, K. and Floyd, S. (1999). A proposal to add

Explicit Congestion Notiﬁcation (ECN) to IP. RFC

2481.

Samaraweera, N. K. G. and Fairhurst, G. (1998). Reinforce-

ment of TCP error recovery for wireless communica-

tion. Computer Communication Review, 28(2).

USING ECN MARKS TO IMPROVE TCP PERFORMANCE OVER LOSSY LINKS

445