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

net/publication/1956855

Improving TCP/IP Performance over Wireless IEEE 802.11 Link

Article · December 2003

Source: arXiv

CITATION READS
1 511

2 authors, including:

Mokhtar Aboelaze
York University
64 PUBLICATIONS 315 CITATIONS

SEE PROFILE

All content following this page was uploaded by Mokhtar Aboelaze on 31 August 2017.

The user has requested enhancement of the downloaded file.

 The Performance TCP/IP over wireless IEEE802.11 Link

Milenko Petrovic Mokhtar Aboelaze

Dept. of Computer Science
York University
4700 Keele St.
Toronto, ON. Canada
M3J 1P3
{petrovi, aboelaze}@cs.yorku.ca

personal digital assistants. Also, these

Abstract applications are no more limited to a run on a PC
Cellular phones, wireless laptops, personal or a workstation that is connected to a local area
portable devices that supports both voice and data network. Wireless terminals are gaining a
access are all examples of communicating devices
widespread acceptance for audio transmission
that uses wireless communication. Sine TCP/IP (and
UDP) is the dominant technology in use in the (cell phones) as well as interactive applications
internet, it is expected that they will be used (and they (web enabled cell phones and PDA devices),
are currently) over wireless connections. In this which access the wired-line network over a
paper, we investigate the performance of the TCP wireless link.
(and UDP) over IEEE802.11 wireless MAC protocol.
The requirements for these different
We investigate the performance of the TCP and UDP
assuming three different traffic patterns. First bulk services are quite diverse. For example ftp like
transmission where the main concern is the applications require a large bandwidth with
throughput. Second real-time audio (using UDP) in error-free transmissions, but there is no
the existence of bulk TCP transmission where the requirement for a tight delay bounds. Voice
main concern is the packet loss for audio traffic. communication on the other hand, can tolerate
Finally web traffic where the main concern is the some data loss without any loss in the perception
response time. We also investigate the effect of using of the voice signal. However, it requires a tight
forward Error Correction (FEC) technique and the end-to-end delay.
MAC sublayer parameters on the throughput and
response time.. TCP/IP has been the dominant network
technology and for good reasons. TCP is a
Key Words: IEEE802.11, Wireless MAC,
reliable transport protocol that is fine-tuned in
TCP performance, UDP performance
order to provide a good performance over
widespread networks with completely different
characteristics. The only problem is that TCP is
1. Introduction tuned for wire-line networks with its very low
error rate. The TCP interprets any packet loss as
There is an ever-increasing demand for a sign of congestion not as a result of corrupted
wireless applications. The current rend in packet due to errors. That leads to TCP trying to
wireless communication is to provide the reduce its transmission rate in order to drain the
infrastructure required to support a wide range of backlog that caused the congestion. When TCP
emerging and existing applications. These is used with wireless networks (with its
applications range from the traditional file relatively higher error rate) that leads to a system
transfer program (ftp), to voice and video working on a reduced capacity since TCP
communication using cellular phones and reduces its transmission rate with each corrupted
packet thinking it is a congestion rather than In this paper we will concentrate on the
error problem. DCF mode of operation. In this mode nodes use
CSMA/CA to compete in order to transmit.
The main question here is “is TCP a good
Nodes do have the option of using Request To
transport protocol for wireless networks,
Send/Clear To Send RTS/CTS in order to avoid
especially the IEEE802.11?” In this paper, we
the hidden terminal problem. On the other hand
investigate the performance of TCP (and UDP)
nodes may decide to use RTS/CTS only for long
over a wireless IEEE802.11 link. Our emphasis
packets or not use it at all.
her is on two parts, the first is how TCP works
under 802.11 in the existence of channel noise. TCP is one of the most successful protocols
The second is the effect of bulk transmission on in the history of computer networks (albeit a
interactive traffic. For interactive traffic we short history). It provides a reliable transport
choose web traffic (TCP) and audio traffic layer connection between two transport layer
(UDP). We used the NS [1] simulator to entities (processes). The TCP protocol adjusts its
simulate different scenarios. transmission rate in order to avoid congestion
and to recover form the congestion when it
The paper is organized as follows: The next
happens. It maintains an estimate for the round
section introduces the IEEE802.11 standard and
trip time, and reacts to packet loss (assuming
briefly explains its salient features. And briefly
that most losses are due to congestion) by
describe the TCP protocol. Section 3 introduces
lowering its transmission rate and slowly
the setup of the network we are simulating and
increases it again in order to fully utilize all the
the three different traffic patterns we used.
available bandwidth between sender and
Section 4 introduces the results of our simulation
receiver.
and discusses them. Section 5 is a conclusion
and future work. TCP uses a dynamic window-sizing
algorithm in which it maintains an estimate of
2. IEEE802.11 and TCP( the current round trip delay. If a packet was not
ACKed within a specific time out (that is a
UDP) function of the round trip time), the packet is
considered lost due congestion and it reduces its
IEEE802.11 [2,3] is a standard that defines transmission rate and probe the network in order
the physical and Medium Access Control to reach the network’s maximum capacity.
sublayer (MAC) of a wireless local area
network. At the physical layer, it defines the Jacobson in [5] proposed two procedures in
physical interface at three different frequency order to improve the performance of TCP. The
bands, 1 and 2 Mbps infrared, 1,2,5.5, and 11 first procedure, known as fast retransmission,
Mbps at the 2.4GHz band using direct sequence which states that if the receiver receives an out
and frequency hopping spread spectrum. It also of order segment, it must send an ACK for the
uses the 5GHz band and Orthogonal FDM to last in-order segment it received and repeats that
achieve a transmission rate up to 54Mbps with each out of order segment that it receives. If
(IEEE802.11a) [4]. At the MAC sublayer, it the sender receives 3 duplicates ACK’s it
defines two modes of operation, the first is assume that that segment is lost and retransmit it
called Distributed Coordination Function (DCF) right away.
that uses an asynchronous, contention based The second procedure, known as fast
access using CSMA/CA (CA stands for collision recovery, which states that in case of fast
avoidance). It also supports the Point retransmission due to 3 duplicate ACK’s, the
Coordination Function (PCF) for a centralized sender should not reduce its window size to 1
contention free transmission for application that segment, but rather cut the current window in
requires a constant bandwidth requirement with half and proceed with the linear increase. A
delay bound. In the DCF mode, it can use more detailed discussion can be found in [6].
Request to send/Clear to send (RTS/CTS) to
minimize the collisions and to deal with hidden Due to the high error rate of the wireless
terminal problems. links, the TCP does not perform in wireless
networks as well as it does in wire-line
networks. Many attempts have been made to We assumed a variation of the Gilbert-
modify TCP in order to accommodate wireless Elliott channel model [11,12], in which the state
links, for a review of some of the methods of the channel alternate between good and bad
proposed, see [7] for a comparison between the states. The duration of the good period and bad
different techniques. period are exponentially distributed random
Improving the TCP throughput on lossy variable with mean τg and τb respectively. In our
wireless links has been studied extensively, In simulation we used τg =0.1 sec. and τb =0.0333
[8] the authors studied segment size adaptation sec. The probabilities of bit-error in each of the
as a way to overcome the losses on the wireless two states are pg and pb respectively. We used
links. In [9] The authors investigated the pg=10-6, and pb= 10-2 similar to [13].
performance of TCP over wireless links, We investigate three different traffic
however they assumed there are no contention at patterns, first, only bulk transmission is allowed
the MAC sublayer, only noise that may result in each node has an infinite amount of data to
error to the frame. The authors in [10] proposed transmit, the main criterion here is the
a proxy-based solution for enhancing the throughput. Next, we assume that we have many
performance of TCP/IP over 802.11b network. web connections and we study the interaction
In this paper, we investigate the between the web traffic (which is more sporadic
performance of the TCP and UDP protocols over than bulk traffic) on the performance. Lastly, we
IEEE802.11 wireless link. We concentrate on assume that there is one bulk connection and
the interactions between the different TCP many voice connections. The voice connection is
streams and its effect on the performance. We assumed to be UDP. The main point here is the
also investigate the access method mechanism bulk throughput and the delay, jitter, and loss
in IEEE802.11 and its effect on the TCP rate of voice packets.
performance. Finally, we investigate FEC based Then, we investigate the effect of Forward
error correction on the performance of the TCP. Error Correction (FEC) on the performance. The
issue of FEC is extremely important in wireless
communication. In bad state, the bit error rate
3. Simulation setup can be as low as 0.01, even with moderate
packet size, that is almost a guaranteed frame
Here, we describe the simulation setup; we error. The FEC can be used to recover from
are assuming a system where more than one errors without the need to retransmit.
mobile node connected to a base station using We assume that the channel state is
IEEE802.11 link with 2Mbps. The base station is estimated at the receiver, and is feedback to the
connected to the wired network through a 5Mbps transmitter. If the channel state is bad, then some
kink with a 2 msec. Delay. The 5 Mbps link is sort of error correction coding is used. This
dedicated to the traffic originating from the base technique is similar to the one proposed in [14],
station. We wanted to study the interaction however they used a different modulation
between the IEEE802.11 and TCP in the technique in order to reduce the bit error rate. In
existence of errors; we avoided any traffic or our simulation, we use a Reed-Solomon error
congestion originating from outside the base correcting code. To implement, this code on our
station. simulation, we calculated the overhead due to
The wireless link is half-duplex and based the parity bits (that basically reduces the
on the AT&T WaveLan network interface card efficiency of the wireless channel to 71% of its
that implements the IEEE802.11 with a 2Mbps minimal value, and we calculated a new bit error
transmission rate and a transmission range of rate after taking the correction into
250 meters. We also assume that mobile nodes at consideration).
the same distance from the base station, and that In the next section, we present results based
each mobile node has a buffer of 50 packets in on simulation using NS [1] and discuss the
its network interface. results and its implication on the performance of
TCP and UDP using IEEE802.11 as a link layer
model.
 4. Simulation Results Window
(in packets) R/F NR/F R/NF NR/NF
Packet size= 100 bytes
In this section, we present the results of 1 56.3 91.8 63.1 80.1
our extensive simulation and discuss them
2 72.35 92.4 79.2 75.8
5 83.9 95.6 70.9 96.2
4.1 Bulk Traffic
10 70.6 97.2 78.8 97.7
First, we investigate transmission of bulk
traffic from one or more connection. The major Packet size = 200 bytes
concern here is the throughput, and the 1 159 175.7 127 167.3
interaction between the retransmission policy of 2 128.2 195.2 158.1 214.1
the TCP and the IEEE802.11. 5 192.4 232.1 184.2 201.1
10 203.1 244.4 153.5 231.9
Table 1 shows the total bandwidth in Kbps
for a 5 bulk TCP connections using four Packet size 500 bytes
different cases, first, R/F we used RTS/CTS for 1 225.8 412 271.1 370.2
every packet transmitted and we used FEC using 2 301.8 421.1 313.9 358.7
Reed-Solomon coding The second, NR/F we did 5 303.9 342.5 341 433.1
not use RTS/CTS and we used FEC for error 10 332.2 485.8 336.8 472.8
correction. The third is R/NF we used RTS/CTS Packet size = 1000 bytes
for every packet and we did not use FEC, the last 1 230.8 585.8 453.6 591.5
one is NR/NF we did not use RTS/CTS and no 2 357 656.8 362.7 599
FEC. We also considered different packet sizes
5 372.6 685.2 343.8 692.6
(100, 200, 500, 1000, and 2000 bytes). The
10 465.4 702.6 227.6 716.4
maximum window size for TCP is 1,2,5, and 10
packets. Packet size = 2000 bytes
1 414.9 610.3 479.8 554.1
From table 1, we can make some important
2 375.4 649.5 621.2 660.4
comments. First, as expected, increasing the
5 498 698.7 490.8 615.9
packet size and increasing the window size
results in a better link utilization. However, one 10 566.1 826 359.7 616
might have expected that using FEC to avoid the
high error rate of the wireless link should have a Table 1: The total throughput of 5 TCP connections
detrimental effect on the system performance. sharing an IEEE802.11 link
Looking at Table 1, with a packet size of 100
bytes, using FEC actually reduces the system
throughput disregarding the use of RTS/CTS Node 1 7.05 53.29 1.25 153.35
While larger packet size (2000 bytes) the use of Node 2 20.54 71.1 157.11 0.31
FEC does improve the system performance.
Node 3 15.09 56.16 0.63 44.84
Another observation is that the use of Node 4 0 89.42 117.29 44.84
RTS/CTS results in decrease in the system
performance. That could be explained by the Node 5 13.6 1.18 214.5 10.35
relatively small cell size in which all the Total 56.3 271.1 490.8 359.7
terminals can hear each other and hence no Packet 100 500 2000 2000
hidden terminal problem. size
Window 1 1 5 10
R/F R/F R/NF R/NF R/NF

Table 2: the throughput of the different nodes and

their packet size and window size
 However, Table 1 does not tell the whole most cases having the combination of no
story, we considered the total throughput that is RTS/CTS and FEC results in the best response
the number of bits per second delivered to the 5 time (except with a window size of 5 where no
nodes. One issue to consider here is the fairness. FEC produces the best response time).
Is the total throughput fairly distributed among
all the nodes? There is no room to put the
detailed results of our experiment. However, Window Size 1 2 5 10
Table 2 shows some of the pathological cases NORTS/NOFEC 0.92 1.01 0.99 0.99
NORTS/FEC 1.06 1.05 0.87 0.89
where there are nodes that are almost shut off RTS/NOFEC 3.92 3.65 3.50 3.87
with a very low throughput, while and the rest of RTS/FEC 4.39 4.27 3.63 3.55
the nodes are sharing the link capacity.
For example in fourth column that describe Table 4: The response time in seconds for 10 nodes,
a system with a 2000 bytes packet, window size web traffic pattern, and 1000 bytes TCP packet size
of 5 packets, and uses RTS/CTS but no error
correction, two nodes are sending/receiving with Table 4 shows the same results as table 3
a rate of 1.25 and 0.63 Kbps, while the other but with a larger packet size (2000 bytes). Here
three nodes are sending with a rate of 157,117, we can see clearly that RTS/CTS only increase
and 214 Kbps. the response time and adding FEC results in
improving the performance for larger window
Most of these cases are suing RTS and CTS. size. But as we see in the next section, although
In our experiment where each node can hear the a larger window size does increase the
rest of the nodes, the use of RTS/CTS degrades throughput of bulk transmission, it might not be
the performance and is not needed. However, in the best policy when we have interactive traffic
other situations where the hidden terminal sharing the same link with the bulk traffic. It is
problem does exist we have to use RTS/CTS and obvious that TCP can recover from the error in
we have to worry about the fairness of the the wireless link quite properly when there is no
protocol. other competition in the wire-line link.

3.2 Bulk and Web Traffic 3.3 Bulk and Audio Traffic
In this section, we assumed 10 web In this part, we consider also one bulk
connections competing for the wireless link. For connection and many voice connections using
the traffic model, we used the web traffic in NS UDP. The emphasis here is on the delay, jitter,
[1]. The main concern here is the response time. and the loss ratio of audio packets. In this case
We have conducted simulations with a variable sine the UDP is not a reliable end-to-end
number of nodes, variable TCP packet length, connection, we had to extensively use the error
and variable TCP window size. Here we present detection/correction. In this paper, we present
the results for packet sizes of 100 and 1000 bytes the results for the packet loss only, for the
only. complete results and discussion, the reader is
referred to [15].

Window Size 1 2 5 10 The voice nodes are transmitting digitized

NORTS/NOFEC 3.85 2.99 1.66 2.68 voice signal. We assume a model where the
NORTS/FEC 2.21 2.51 3.99 1.78 speaker alternate between talk spurts and silence.
RTS/NOFEC 6.33 4.98 9.25 5.44 The talk spurts and silence are assumed to be
RTS/FEC 10.53 5.23 9.63 4.60 exponentially distributed with average of 1.0 and
1.35 sec. respectively [16] During the talk spurt,
Table 3: The response time in sec. for 10 nodes, web we assume that digitized voice at 32Kbps is
traffic pattern, and 100 bytes TCP packet size being generated.
Table 3 shows the response time for 10 For real-time voice transmission, the
nodes using web traffic pattern for different important performance measure is the end-to-
window size and different combination of FEC end delay, the rate of lost packets, and the jitter.
and RTS/CTS. From Table 3 we can see that in Studies indicate that 1-2% loss of audio data
doesn’t have an impact on the perception of the window size for the background TCP connection.
audio signal. Besides the lost packets, packets A small UDP packet size decreases the
that reach the destination delayed may not be efficiency since the UDP and IP headers are
useful anymore and will be thrown away by the constant and don’t depend on the UDP packet
receiver and will have the same effect as the lost size. Also decreasing the window size of the
packets. background bulk TCP connection will definitely
reduce the bulk traffic throughput as we saw in
Table 5 shows the probability of packet loss
Table 1. However that could be the best way to
vs. the UDP packet size and the window size of
share the medium between bulk TCP
the TCP source of a 7 bi-directional voice
transmission and real-time voice signals.
connections. For the TCP source, we assumed a
packet size of 1500 bytes, and a variable window
size. The packet loss in this model is due to 2
reasons. First, a packet may be lost because it 5. Conclusion
has exhausted the number of retransmission by
the IEEE802.11 and thus thrown away, or In this paper we presented an analysis of
because it reached the destination late (more the performance of TCP and UDP on a wireless
than 0.5 sec. delay) and thus deemed unusable link that is connected to a wire-line network. We
by the destination and thrown away, the data in considered three types of traffic, bulk traffic
Table 5 is for the total loss rate. The first row in where each node has an infinite amount of data
the table indicates the UDP packet size, the first to transmit and the total throughput and the
column indicates the TCP window size in fairness is the dominant performance measure.
packets for the background traffic (note that the Web traffic with requests from the client and
TCP packet size is fixed at 1500 bytes). response from the server, where the response
time is the dominant performance measure.
Since voice transmission is considered O.K. Finally Real-time voice calls in the existence of
if the probability of lost packets is 1-2%, from a background connection of bulk transmission
Table 5, we can see that voice transmission where the loss rate for voice packets is the
using IEEE802.11 DCF is possible only if we dominant performance measure.
used a small packet size (the error is the major
problem for larger packet sizes). Also, a small
TCP window size is helpful however; even for
large window sizes the loss rate is still
References
acceptable for small UDP packet size.
[1] VINT project U.C. Berkely/LBNL, bs2:
W UDP Packet Size network simulator, http://www-
200 500 600 800 1000 mach.cs.berkeley.edu/ns 2001
1 0.2% 0.40% 0.2% 0.33% 8.5%
2 0.95% 0.36% 0.28% 0.28% 7.2%
[2] IEEE Std 802.11-1997 Standard for
4 0.45% 0.46% 0.29% 0.37% 7.6% Wireless LAN Medium Access Control
8 1.19% 0.36% 0.46% 0.43% 7.2% (MAC) And Physical Layer (PHY)
16 1.5% 0.23% 0.41% 0.45% 7.2% Specifications, 1997.
32 1.2% 0.5% 0.38% 0.5% 8.2%
[3] IEEE Std 802.11b-1999/Cor 1-2001
Standard for wireless LAN medium access
Table 5: The probability of loss of 7 audio control (MAC) and physical layer (PHY)
connections using UDP in the presence of 1 bulk TCP
specifications 2001
connection
[4] IEEE Std 802.11a-1999 , 30 Dec. 1999
From our experiment it is almost impossible
Supplement to IEEE standard wireless LAN
to have 7 voice connections with a meaningful
Medium Access Control (MAC) and
loss rate with a UDP packet size more than 800
Physical Layer (PHY)
bytes.
[5] V. Jackobson “Modified TCP Congestion
The best results are achieved with a small
Avoidence Algorithm” end2end –interest
UDP packet size (200 bytes) and a small
 mailing list, April 1990 Also available at [15] M. Petrovic, “The Performance of TCP/IP
ftp://ftp.ee.lbl.gov/email/vanj.90apr30.txt in Wireless Environment using IEEE802.11
MAC” Master Thesis, Dept .of Computer
[6] W. R. Stevens “TCP/IP Illustrated, Volume
Science, York University, August 2002.
1”Adison-Wesley, 1994
[16] J. Gruber and L. Strawczynski “Subjective
[7] H. Balakrishnan, V.N. Padmanabhan, S.
Effects of Variable Delay and Speech
Seshan,, and R. H. Katz. “A comparison of
Clipping in a Dynamically Managed Voice
mechanisms for improving TCP
System” IEEE Transactions on
performance over wireless links”,
Communication, 1985 COM-33 (8) pp801-
IEEE/ACM Transactions on Networking,
808
Volume: 5 Issue: 6, Dec. 1997 Page(s): 756
–769
[8] H. Kim, and C. Han “Reducing TCP
Response time in Face of Wireless Uplink
Losses”, IEEE International Conference on
Communication pp 262-266
[9] H.-K. Shiu, Y.-H. Chang, T.-C Hou, and
C.-S. Wu, “Performance Analysis of TCP
over Wireless Link with Dedicated Buffers
and Link Level Error Control”,IEEE
International Conference on
Communication pp3211-3216, 2001
[10] L. Munoz, M. Garcia, J. Choque, and R.
Aguero, “Optimizing Internet Flows over
IEEE802.11b Wirelss Local Area Networks:
A Peformance-Enhancing Proxy Based on
Forward Error Correction”, IEEE
Communication magazine, pp 60-67 Dec.
2001
[11] E. N. Gilbert, “Capacity of a Burst-noise
Channel”, Bell System Technical Journal
vol. 39, pp 1253-1265, Sept. 1960.
[12] E. O. Elliot, Estimates of Error Rates for
Codes on Burst-noise Channel”, Bell
System Technical Journal, vol 42, pp1977-
1997, Sept. 1963
[13] J. G. Kim, and M. M. Krunz, “Bandwidth
Allocation in Wireless Networks with
Guaranteed packet-Loss Peformance”
IEEE/ACM Transactions on Networking,
Vol. 8, No. 3, pp 337348 June 2000
[14] V. K. N. Lau; Yu-Kwong Kwok ,
“CHARISMA: a novel channel-adaptive
TDMA-based multiple access control
protocol for integrated wireless voice and
data services” Wireless Communications
and Networking Conference WCNC. 2000
Page(s): 507 -511 vol.2, 2000

View publication stats