0% found this document useful (0 votes)
76 views21 pages

TCP Congestion Control Over 3G Communication Syste

This document analyzes the performance of TCP congestion control algorithms over 3G networks through experimental evaluation. The authors conducted measurements over a live UMTS network involving around 3000 flows and more than 40 hours, comparing New Reno, BIC and Westwood+. Results show that a single downlink flow is far from saturating the channel bandwidth, and the considered TCP stacks provide similar results. The 90th percentile goodput of a single downlink flow is less than 230 kbps compared to the 384 kbps UMTS downlink capacity.

Uploaded by

Cabdalla
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
76 views21 pages

TCP Congestion Control Over 3G Communication Syste

This document analyzes the performance of TCP congestion control algorithms over 3G networks through experimental evaluation. The authors conducted measurements over a live UMTS network involving around 3000 flows and more than 40 hours, comparing New Reno, BIC and Westwood+. Results show that a single downlink flow is far from saturating the channel bandwidth, and the considered TCP stacks provide similar results. The 90th percentile goodput of a single downlink flow is less than 230 kbps compared to the 384 kbps UMTS downlink capacity.

Uploaded by

Cabdalla
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 21

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220876639

TCP Congestion Control over 3G Communication Systems: An Experimental


Evaluation of New Reno, BIC and Westwood+

Conference Paper · September 2007


DOI: 10.1007/978-3-540-74833-5_7 · Source: DBLP

CITATIONS READS

6 129

2 authors:

Luca De Cicco Saverio Mascolo


Politecnico di Bari Politecnico di Bari
64 PUBLICATIONS   1,607 CITATIONS    111 PUBLICATIONS   4,025 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

F/050136/01/X32 View project

Modeling and Control of Adaptive Video Streaming Deliver Systems View project

All content following this page was uploaded by Luca De Cicco on 28 May 2014.

The user has requested enhancement of the downloaded file.


TCP Congestion Control over 3G Communication
Systems: an Experimental Evaluation of New Reno,
BIC and Westwood+
Luca De Cicco and Saverio Mascolo

{ldecicco,mascolo}@poliba.it

Dipartimento di Elettrotecnica ed Elettronica, Politenico di Bari, Via Re David 200, Italy

Abstract. One of TCP's key tasks is to react and avoid network congestion episodes which nor-
mally arise in packet switched networks. A wide literature is available concerning the behaviour
of congestion control algorithms in many dierent scenarios and several congestion control algo-
rithms have been proposed in order to improve performances in specic scenarios. In this paper
we focus on the UMTS wireless scenario and we report a campaign of measurements that in-
volved around 3000 ows and more than 40 hours of measurements using three dierent TCP
stacks: TCP NewReno, which is the congestion control algorithm standardized by IETF, TCP
BIC which is the default congestion control algorithm adopted by the Linux operating system,
and TCP Westwood+ also available in the Linux kernel. The experimental evaluation has been
carried out by accessing the public Internet using an UMTS card. Distribution of measurements
of goodputs, RTTs over time, packet loss ratios, number of timeouts and Jain Fairness Indices are
reported through cumulative distribution functions. Moreover, the eciency of each TCP version
in transferring les has been evaluated by varying the le size in the range from 50 KB up to
500 KB . The cumulative distribution functions reported in the paper show interesting results: 1)

a single downlink ow is far from saturating the channel bandwidth; 2) considered TCP stacks
provide similar results; 3) 90th (50th) percentile of the goodput of a single downlink ow is less
or equal then 230 kbps (120 kbps) compared to a nominal 384 kbps UMTS downlink channel.

TCP, congestion control, 3G, UMTS


TR 17.04.07 L. De Cicco, S. Mascolo

1 Introduction

Since 2001, the year that the rst commercial UMTS network was deployed by NTT
DoCoMo in Japan, many telecom operators have launched UMTS access to subscribers.
As many 3G networks are emerging, it is important to evaluate how dierent TCP

congestion control mechanisms behave in such networks. The UMTS network provides

wide area Internet wireless access with downlink speeds up to 384 kbps and round trip

times in the order of 300 ms thus providing a viable solution for multimedia and for

VoIP applications.

It is known that eciency of TCP as a transport protocol degrades when lossy links

are present in the routing path [1][6], such as in the case of wireless links. For this

reason, the UMTS link layer implements the Radio Link Control (RLC) protocol that

masks the lossy channel to upper layers through retransmissions. In this way, in-order

packet delivery and loss probability less than 1% are provided [3]. The reliability of

the link layer comes at the cost of an highly variable segments delay as seen at the

transport layer when frames are retransmitted at the link layer, thus possibly leading

to spurious timeouts [11]. The impact of spurious timeouts has been studied extensively

so far and it has been often considered as one of the major causes of TCP throughput

degradation. However, against this common belief authors in [15] have found that, in

the case of a well-designed UMTS network and in the static scenario, the number of

spurious timeouts is very low, thus having negligible impact on TCP throughput.

Another issue raised by the UMTS link layer is the variability of the available band-

width that is caused by the channel state scheduling. In [3] authors address both rate

and delay variabilities, showing the negative eects of delay variability on throughput

achieved by TCP.

In this paper we present the results obtained through an extensive campaign of

measurements obtained over a live UMTS network. Both downlink and uplink mea-

2
TR 17.04.07 L. De Cicco, S. Mascolo

surements of goodputs, RTTs, queuing times, packet loss ratios, number of timeouts

and Jain Fairness Indices have been collected. We have considered three dierent TCP

stacks that are available in the Linux kernel: TCP NewReno [7], which is the congestion

control algorithm standardized by IETF, TCP BIC [16], which is the default congestion

control algorithm adopted by the Linux operating system, and TCP Westwood+, which

has been proposed in [8,12] and is also available in the Linux kernel.

The rest of the paper is organized as follows: in Section 2 we summarize the prior

work on live UMTS measurements and we briey describe the considered congestion

control algorithms; in Section 3 we describe the considered experimental testbed; Sec-

tion 4 reports experimental results whereas a discussion of the results is reported in

Section 5. Finally, Section 6, draws the conclusions.

2 Related Work

2.1 Live UMTS network performance evaluation

The academic literature contains a plethora of simulation studies about TCP perfor-

mances over UMTS and GPRS, but very few papers have addressed the performance

evaluation of a live UMTS network.

In [10] authors report results obtained by an experimental investigation carried out

in an early deployment of two 3G networks in near-ideal conditions, i.e. no other user

was accessing the network. In the paper it is noticed that the UMTS link is aected

by very few spurious timeouts and that the employment of the Eifel algorithm did not

provide any throughput improvement.

In [13] authors report goodput measurements obtained by accessing the public

3G/UMTS network to transfer les with dierent size; the paper focuses only on the

downlink and does not report any data regarding RTT variability, typical number of

timeouts and fairness indices.

3
TR 17.04.07 L. De Cicco, S. Mascolo

Authors of [2] provide an IP and TCP level measurement of the UMTS downlink

and uplink in both static and mobile scenario. The paper focuses on nding the optimal

settings for the MSS and the initial receiver window; throughput measurements are

given for a limited number of considered scenarios.

2.2 TCP Congestion control algorithms

One of the most important tasks that the TCP addresses is regulating the sending rate

in order to avoid the network congestion. In [9] Van Jacobson has proposed a solution

to the network congestion control problem that mainly consists of two distinct phases:

a probing phase and a decreasing phase (the well-known AIMD paradigm [4]). During

the probing phase the link capacity is probed using an exponential growth law which is

called slow start, or a linear growth law that is called congestion avoidance. The con-

gestion control algorithm switches from the probing phase to the decrease phase when

a three duplicate acknowledgment (3DUPACK) or a timeout is experienced indicating

that a congestion event has taken place. During the decreasing phase the congestion

window cwnd is multiplicatively decreased in reply to the congestion episode. A pseudo-


code of the TCP congestion control behaviour is the following:

 When an ACK is received: cwnd ← cwnd + a

 When a 3 DUPACKs are received: cwnd ← cwnd − b · cwnd

During the congestion avoidance phase of TCP NewReno a is set equal to 1/cwnd,
meaning that the congestion window is increased by one packet for each RT T , whereas
in the slow start phase a is set to 1 meaning that the cwnd is doubled each RT T . The

constant b is set to 0.5 so that the cwnd is halved after a congestion episode. When

a timeout is experienced then, the cwnd is set to 1 segment and the slow start phase

takes place.

4
TR 17.04.07 L. De Cicco, S. Mascolo

TCP BIC [16] congestion control is made of two parts: the binary search increase

phase and the additive increase phase. In the binary search phase the congestion window

setting is performed as a binary search problem. After a packet loss, the congestion

window is reduced by a constant factor b, cwndmax is set to the window size before the

lost and cwndmin is set to the value of congestion window after the loss (cwndmin =
b · cwndmax ). If the dierence between the congestion window middle point (cwndmax +
cwndmin )/2 and the minimum congestion window is lower than a threshold Smax the

protocol starts a binary search algorithm increasing the congestion window to the middle

point, otherwise the protocol enters the linear increase phase and increments the

congestion window by one for each received ACK. If BIC does not get a loss indication

at this window size, then the actual window size becomes the new minimum window;

otherwise, if it gets a packet loss, the actual window size becomes the new maximum.

The process goes on until the window increment becomes lower than the Smin threshold

and the congestion window is set to cwndmax . If the window grows more than cwndmax ,
the protocol enters into a new phase (max probing) that is specular to the previous

phase; that is, it uses the inverse of the binary search phase rst and then the additive

increase.

TCP Westwood+ [12,8] is a sender side modication of TCP NewReno in which

the multiplicative decreasing phase is replaced with an adaptive decreasing phase. In

particular, after a congestion episode cwnd is set such that the bandwidth which is

available at the time of congestion is exactly matched. The available bandwidth is

estimated by counting and averaging the stream of ACK packets. In particular, when

three DUPACKs are received, the congestion window cwnd is set equal to the estimated
bandwidth (BW E ) times the minimum measured round trip time (RT Tmin ). After a

timeout the slow start threshold is set equal to BW E · RT Tmin and the cwnd is set to

one.

5
TR 17.04.07 L. De Cicco, S. Mascolo

3 Experimental Testbed

In order to carry out our experiments we have set up a machine at the FTW research

center in Vienna, installing a Linux Kernel with Web100 support. The TCP ows have

been generated and received using a modied version of iperf [14] which uses the

libnetmeas [5] library that we have developed in order to automatically get instanta-

neous values of internal kernel variables such as cwnd, RTT, ssthresh, timeouts.

We have found out that the telecom operator uses a transparent proxy that probably

implement some sort of Split Stream solution, so that when the user downloads the le

for the rst time, the le is cached on a proxy that is located in the telecom operator

network. This way the second download provides better results, since the path results

shorter than that of the rst download. For this reason we have not used an HTTP

server as the sender and an HTTP client (such as wget) for the receiver.

The scenarios and the testbed that we have used in this investigation are aimed

at characterizing both the uplink and downlink UMTS channels in the most common

user scenarios. The user equipment (UE) is a Nokia 6630 mobile phone connected to a

laptop via USB 2.0 that was located at the C3Lab, Politecnico di Bari (Italy) accessing

the public UMTS network using a commercial card provided by a local mobile operator

(Figure 1).

Fig. 1. Experimental testbed used in our evaluation

6
TR 17.04.07 L. De Cicco, S. Mascolo

The UE was static and was accessing the UMTS network in an indoor environment

so that handovers could not occur during measurements. The nominal value declared

by the telecom operator for the downlink (uplink) channel is 384 kbps (64 kbps).

Based on the results obtained in [2] we have xed the maximum segment size to 1500

bytes. Moreover we have set the initial receiver advertisement window to the default

value of 64 Kb which is well above the bandwidth delay product so that we are sure

that the bottleneck is located in the UMTS network.

For each connection we have collected a very rich set of measurement including

goodputs, RTTs, queuing times, number of timeouts, retransmission ratios. We have

evaluated the Jain Fairness Index as follows [4]:

P 2
N
i=1 x i
JF I = PN 2
N i=1 xi

where xi is the mean goodput obtained by the i-th ow accessing the downlink or the

uplink.

We have evaluated TCP congestion control algorithms in the following scenarios:

1. one ow over the UMTS downlink or uplink;

2. two or four homogeneous ows sharing the UMTS downlink or uplink;

3. short le transfers of 50KB, 100KB, 200KB, 500KB, les on both UMTS downlink

and uplink;

For each scenario we have run experiments using the Linux Kernel implementation of

TCP Bic, TCP NewReno and TCP Westwood+[12,8]. All tests, except those for the

short le transfer scenario, lasted approximately 100 seconds each, thus counting for

more than 40 hours of active measurements of downlink and uplink UMTS channel.

7
TR 17.04.07 L. De Cicco, S. Mascolo

In order to perform a fair comparison, we have run tests by rotating TCP congestion

control algorithms and scenarios and repeating tests in dierent days and in dierent

hours of the day.

4 Experimental Results

We have performed an extensive experimental evaluation of the three congestion control

algorithms by collecting measurements of around three thousand ows for hours of

active measurements. In this section we report the results collected for both downlink

and uplink.

4.1 Goodput, link utilization and fairness

The case of downlink ows


Figures 2 (a), (b) and (c) report the cumulative distribution functions for the re-

sults obtained for the cases of one, two and four ows respectively, sharing the UMTS

downlink, whereas mean goodputs and standard deviations are summarized in Table 1.

It is noteworthy that the three congestion control protocols provide similar results

in all the cases, the only remarkable dierence being the case of the single ow (Figure

2 (a)) where TCP NewReno obtain 12% and 8% goodput less than TCP Westwood+

and TCP Bic. The median values in the case of one, two and four ows are in the ranges

[101, 121], [76, 87] and [71, 78] kbps respectively, whereas the upper 10th percentile ex-

periences bandwidth in the ranges [220, 249], [128, 146] and [121, 139]kbps, respectively.
Figure 2 (d) shows the downlink utilization that is computed by averaging goodputs

of all the experiments performed for each considered protocol when the number of

ows sharing the link are one, two or four. It is worth noting that all the three tested

TCP variants provide less than 40% of bandwidth utilization in the single ow case.By

increasing the number of ows sharing the bottleneck the utilization reaches about 90%

8
TR 17.04.07 L. De Cicco, S. Mascolo

(a) (b)

(c) (d)

Fig. 2. Cumulative distribution function of the goodput measured in the case of 1 ow
(a), 2 ows (b), 4 ows (c) sharing the UMTS downlink; bandwidth utilization (d)

New Reno BIC Westwood+

#Flows E[x] σ(x) Ch. Utiliz. E[x] σ(x) Ch. Utiliz. E[x] σ(x) Ch. Utiliz.

1 122.03 65.38 31.8% 133.98 69.41 34.9% 138.58 77.10 36.1%


2 86.08 36.20 44.8% 85.82 40.51 44.7% 90.41 46.12 47.1%
4 80.13 38.53 83.5% 83.83 30.68 87.3% 82.46 40.60 85.9%
Table 1. Average and standard deviation values (in kbps) of goodput for the UMTS
downlink channel

9
TR 17.04.07 L. De Cicco, S. Mascolo

in the four ows case. We will return later on this topic in Section 5 where we will

discuss possible reasons of the very low link utilization in the one ow case.

Finally, it is worth to notice that goodputs reported by Catalan et al. [2] in the

single ow case averaged over three experiments are similar to our results.

Moreover, concerning the fairness indices we have found that the considered TCP

variants provide similar average values of the JF I both in the two and four ows cases.

We have obtained values around 0.94 for the two ows case and around 0.86 for the

four ows case.

The case of uplink ows


The evaluation of the UMTS uplink channel have led to the results depicted in Figure

3 (a). Also in the uplink scenarios the goodputs obtained by the considered TCP con-

gestion control algorithms are very similar, TCP Bic performing slightly better in the

single ow case.

In the case of uplink ows Table 2 enlights that the values of the standard deviation

of goodput is very low for all the considered TCP stacks and in all cases. The median

values in the case of one, two and four ows are respectively in the ranges [53, 55],
[24, 26] and [14, 15] kbps, whereas the upper 10th percentile experiences bandwidths

that are respectively in the ranges [57, 59], [31, 33] and [17, 20]kbps.

Dierently from the case of downlink ows, the protocols that we have tested have

provided nearly full uplink utilization even in the single ow case as it is shown in

Figure 3 (d).

In the uplink scenario Jain Fairness Index is very high for the considered TCP

stacks, being around 0.99 for the two ows case and around 0.95 for the four ows case.

10
TR 17.04.07 L. De Cicco, S. Mascolo

(a) (b)

Fig. 3. Goodput cumulative distribution function for the uplink scenario (a): 1 ow in
solid lines , 2 ows dashed lines and 4 ows in dot-dashed lines. Link utilization (b)

New Reno BIC Westwood+

#Flows E[x] σ(x) Ch. Utiliz. E[x] σ(x) Ch. Utiliz. E[x] σ(x) Ch. Utiliz.

1 56.44 2.09 88.2% 57.51 2.11 89.9% 55.62 1.74 86.9%


2 29.40 3.95 91.9% 29.44 2.31 92.0% 29.32 2.19 91.6%
4 14.78 4.08 92.4% 14.82 2.81 92.6% 14.72 2.51 92.02%
Table 2. Average and standard deviation values (in kbps) of goodput for the UMTS
uplink channel

11
TR 17.04.07 L. De Cicco, S. Mascolo

4.2 RTT and queuing time

The case of downlink ows.


Figures 4 (a), (b) and (c) show the cumulative distribution functions of the round trip

1
times (RTT) and the queuing times, hereinafter named , in the case of the downlink

channel, while Table 3 collects mean values, RTT standard deviations and average

queuing times for the considered scenarios.

Even though the minimum RTT experienced in our evaluation is around 300 ms,
the measured queuing times are very high and they do not depend on the number of

ows, thus suggesting that in the single ow case we have an excessive queuing. TCP

NewReno exhibits a slightly inated value of RT T quantied in 100 ms in the one ow

case, due to larger queuing time. In all evaluated scenarios, TCP Westwood+ provides

less queuing time with respect to the other two algorithms. This result is coherent with

the Westwood+ unique feature of clearing all queuing after a congestion.

New Reno BIC Westwood+

#Flows E[RT T ] σ(RT T ) E[tq ] E[RT T ] σ(RT T ) E[tq ] E[RT T ] σ(RT T ) E[tq ]
1 1550 1096 1248 1457 897 1137 1469 1110 1125
2 1297 624 953 1369 691 1024 1219 508 873
4 1338 488 995 1159 230 825 1102 221 765
Table 3. Average and standard deviation values (in ms) of RTT for the UMTS downlink
channel. Average values (in ms) of queuing.

The case of uplink ows


Figures 5 (a), (b) and (c) show the round trip time and the queuing time cumulative

distribution functions in the case of uplink ows while Table 4 collects the values of mean

RTT, RTT standard deviation and average queuing time. Similarly to the downlink

1
The round trip time is dened as sum of a propagation time, which can be evaluated as the minimum RTT,
and the queuing time tq .

12
TR 17.04.07 L. De Cicco, S. Mascolo

(a) (b)

(c)

Fig. 4. RTT (solid lines) and queuing time (dashed lines) cumulative distribution func-
tion in the case of 1 ow (a), 2 ows (b) and 4 ows (c) sharing the UMTS downlink.

13
TR 17.04.07 L. De Cicco, S. Mascolo

scenario, TCP Westwood+ exhibits less queuing time except the case of four ows

where TCP NewReno provides slightly less queuing time. Dierently from the case

of downlink ows, the RTT and queuing time tends to increase with the number of

concurrent ows.

(a) (b)

(c)

Fig. 5. RTT (solid lines) and queuing time (dashed lines) cumulative distribution func-
tion in the case of 1 ow (a), 2 ows (b) and 4 ows (c) sharing the UMTS uplink.

14
TR 17.04.07 L. De Cicco, S. Mascolo

New Reno BIC Westwood+

#Flows E[RT T ] σ(RT T ) E[tq ] E[RT T ] σ(RT T ) E[tq ] E[RT T ] σ(RT T ) E[tq ]
1 1471 113.2 927 1452 113.2 904 1368 87.1 828
2 1887 80.2 1416 1998 104.3 1521 1869 79.5 1364
4 2222 72.7 1556 2276 70.7 1621 2301 78.7 1587
Table 4. Average and standard deviation values (in ms) of RTT for the UMTS uplink
channel. Average values (in ms) of queuing.

4.3 Timeouts and packet retransmission percentage

Table 5 summarizes the average number of timeouts obtained on the UMTS downlink

and uplink in the considered scenarios. The results obtained in the downlink and uplink

scenarios are very dierent and can be the source of the very dierent link utilization

that we have reported in Section 4.1.

In the downlink case the ows suer around ve timeouts throughout the 100 s
duration of the connection regardless of the number of ows and the congestion control

algorithm used. On the other hand, in the uplink ows case, the number of timeouts is

negligible.

New Reno BIC Westwood+

#Flows Downlink Uplink Downlink Uplink Downlink Uplink

1 5.84 0.15 4.82 0.13 5.27 0.15


2 6.44 0.04 5.43 0.03 6.20 0.06
4 5.35 0.45 5.24 0.35 5.80 0.43
Table 5. Average number of timeouts for downlink and uplink (the duration of the
connection is 100 s)

Due to space limitation we can't show the cumulative distribution function of packet

loss ratios, but we report the mean value and the standard deviation in the Table 6. By

considering the values reported in the table we can observe that there is no remarkable

15
TR 17.04.07 L. De Cicco, S. Mascolo

dierence in the considered TCP congestion control algorithms. In the case of downlink

ows, the retransmission percentage increases with the number of ows and it is below

11%, whereas in the case of uplink ows the fraction of retransmitted packets is less

than 1%.

New Reno BIC Westwood+

Downlink Uplink Downlink Uplink Downlink Uplink

#Flows E[p] σ(p) E[p] σ(p) E[p] σ(p) E[p] σ(p) E[p] σ(p) E[p] σ(p)
1 7.13 4.89 0.04 0.12 6.75 2.97 0.03 0.10 7.79 6.05 0.03 0.10
2 7.96 5.28 0.02 0.12 7.97 4.41 0.02 0.12 10.12 6.86 0.02 0.10
4 10.95 5.85 0.40 0.98 10.05 3.77 0.23 0.52 11.69 4.57 0.27 0.55
Table 6. Average retransmission percentage (%) and standard deviation in the case of
uplink and downlink.

4.4 Goodput versus le size

In this section we investigate the impact of the le size on the goodput of TCP. We

have collected goodput measurements for le size in the range from 50 KB to 500 KB ,
in order to nd out if the slow start phase degrades goodput in the case of small size

le transfers.

Figures 6 (a) and (b) show results respectively for the downlink and uplink cases.

In the case of downlink ows, all considered TCP variants provide similar results and

the goodput is essentially constant when the le size increases, showing a maximum at

100 KB in the case of TCP NewReno and TCP Westwood+. Also in the case of uplink

ows the goodput is constant, slightly increasing of 10 kbps when the le size grows

from 50 KB to 500 KB .

Thus, we can conclude that the slow start phase does not cause remarkable eects

on the goodput in both the downlink and uplink channels.

16
TR 17.04.07 L. De Cicco, S. Mascolo

(a) (b)

Fig. 6. Goodput vs le size for (a) downlink and (b) uplink channel

5 Discussion of results

In the previous sections we have reported the results of our extensive UMTS evaluation

and we have found that the performance of the considered TCP congestion control

algorithms are comparable in all the scenarios we have evaluated.

Furthermore, we have found that in the case of uplink ows there are no remarkable

issues, since all TCPs provide satisfactory channel utilization and exhibit very low

number of timeouts. On the contrary, Section 4.1 has shown that in the case of the

downlink scenario, all TCPs provided a very low utilization of UMTS links in the single

ow case.

One could argue that the cause of the poor utilization could be due to congestion

somewhere on the path or on the UMTS downlink channel, but the average channel

utilization increases with the number of ows as it can be noticed by looking at Table 1.

To gain an insight into the reason of the poor link utilization let's consider the number

of timeouts in the case of the downlink ows as summarized in Table 5. It can be seen

that the number of timeouts is roughly constant and does not depend on the number of

17
TR 17.04.07 L. De Cicco, S. Mascolo

ows. This observation suggests that timeouts are not due to congestion, otherwise we

should expect to measure more timeouts in the case of multiple ows over the downlink.

Thus, we argue that the reason of the poor downlink utilization in the case of single

ow is mostly due to the high number of timeouts that impose an upper bound to the

achievable goodput.

In fact, the inated RTT values due to large buering, which has been already

discussed in Section 4.2, make the retransmission timeout longer, thus implying a very

long time spent in recovering timeout events [11]. Using trivial arguments, the average

time tout spent resolving timeouts is the product of the average number of timeouts

Ntout that aect the connection and the average retransmission timeout value T0 :

E[ttout ] = E[Ntout ] · E[T0 ] ∼


= 35 s

which in our case, by considering that the connections last 100 s, results in 35% of the

connection time. It is worth to notice that this value matches the link utilization shown

in Section 4.1 (see Figure 2 (d)).

6 Conclusions

In this paper an extensive TCP performance evaluation over a live UMTS network

by measuring both downlink and uplink performance indices for three dierent TCP

congestion control algorithms is reported. The main ndings can be summarized as

follows: (i) the considered TCP congestion control algorithms performed similarly both

in downlink and uplink scenarios; (ii) the UMTS uplink channel did not exhibit any

remarkable issues, providing good channel utilization and very low number of timeouts

and packet retransmissions; (iii) a very high number of timeouts has been observed in

our measurements in the case of downlink channel that does not seem to be caused by

18
TR 17.04.07 L. De Cicco, S. Mascolo

congestion; (iv) the UMTS downlink channel utilization is poor in the single ow case

because of the joint eect of the very high number of retransmissions and the inated

RTT due to queuing.

7 Acknowledgments

We would like to thank Fabio Ricciato for allowing us to use a server at FTW. This work

has been partially supported by the MIUR-PRIN project no. 2005093971 "FAMOUS

Fluid Analytical Models Of aUtonomic Systems" and by Financial Tradeware S.r.l..

References
1. H. Balakrishnan, VN Padmanabhan, S. Seshan, and RH Katz. A comparison of mechanisms for improving
TCP performance overwireless links. Networking, IEEE/ACM Transactions on , 5(6):756769, 1997.
2. M. Catalan, C. Gomez, D. Viamonte, J. Paradells, A. Calveras, and F. Barcelo. TCP/IP analysis and opti-
mization over a precommercial live UMTS network. Wireless Communications and Networking Conference,
2005 IEEE , 3, 2005.
3. M.C. Chan and R. Ramjee. TCP/IP Performance over 3G Wireless Links with Rate and Delay Variation.
Wireless Networks , 11(1):8197, 2005.
4. D.M. Chiu and R. Jain. Analysis of the increase and decrease algorithms for congestion avoidance in
computer networks. Computer Networks and ISDN Systems , 17(1):114, 1989.
5. L. De Cicco. Libnetmeas, 2006.
6. David A. Eckhardt and Peter Steenkiste. Improving wireless lan performance via adaptive local error
control. In ICNP , pages 327338, 1998.
7. S. Floyd and T. Henderson. RFC2582: The NewReno Modication to TCP's Fast Recovery Algorithm.
Internet RFCs , 1999.
8. L.A. Grieco and S. Mascolo. Performance evaluation and comparison of Westwood+, New Reno, and Vegas
TCP congestion control. ACM SIGCOMM Computer Communication Review , 34(2):2538, 2004.
9. V. Jacobson. Congestion avoidance and control. ACM SIGCOMM Computer Communication Review ,
25(1):157187, 1995.
10. M. Kohlwes, J. Riihijarvi, P. Mahonen, F. GmbH, and G. Dusseldorf. Measurements of TCP performance
over UMTS networks in near-ideal conditions. Vehicular Technology Conference, 2005. VTC 2005-Spring.

2005 IEEE 61st , 4.

19
TR 17.04.07 L. De Cicco, S. Mascolo

11. R. Ludwig, B. Rathonyi, A. Konrad, K. Oden, and A. Joseph. Multi-layer tracing of TCP over a reliable
wireless link. Proceedings of the 1999 ACM SIGMETRICS international conference on Measurement and

modeling of computer systems , pages 144154, 1999.


12. S. Mascolo, C. Casetti, M. Gerla, MY Sanadidi, and R. Wang. TCP westwood: Bandwidth estimation for
enhanced transport over wireless links. Proceedings of the 7th annual international conference on Mobile

computing and networking , pages 287297, 2001.


13. K. Pentikousis, M. Palola, M. Jurvansuu, and P. Perala. Active goodput measurements from a public
3G/UMTS network. Communications Letters, IEEE , 9(9):802804, 2005.
14. A. Tirumala and J. Ferguson. Iperf, 2001.
15. F. Vacirca, T. Ziegler, and E. Hasenleithner. An algorithm to detect TCP spurious timeouts and its
application to operational UMTS/GPRS networks. Computer Networks , 50(16):29813001, 2006.
16. L. Xu, K. Harfoush, and I. Rhee. Binary increase congestion control (BIC) for fast long-distance networks.
INFOCOM 2004. Twenty-third AnnualJoint Conference of the IEEE Computer and Communications So-

cieties , 4, 2004.

20

View publication stats

You might also like