A Split Connection TCP Proxy in LTE Networks

Viktor Farkas, Balázs Héder, and Szabolcs Nováczki

Nokia Siemens Networks

{viktor.farkas,balazs.heder,szabolcs.novaczki}@nsn.com

Abstract. High quality and ubiquitous Internet access is a key feature

of today’s mobile systems such as LTE. While LTE can provide com-
petitive peak data rates and a relatively low latency, there is still room
for solutions improving end-users’ Quality-of-Experience by optimizing
services running over the LTE infrastructure. Being the most widespread
transport protocol, TCP is in the main focus of such research projects.
A widely recommended solution for TCP performance improvement is
the split connection TCP proxy that divides the end-to-end TCP connec-
tion into two independent connections, that results increased throughput
and faster error recovery. This paper investigates the performance of a
split connection TCP proxy deployed in LTE’s SAE-GW. Numerical re-
sults show significant performance improvement of file downloading, web
browsing and video steaming applications in case of not congested trans-
port networks.

Keywords: LTE, PEP, TCP proxy, split TCP.

1 Introduction

Latest advances in cellular mobile technology enables users to access most of

their favorite services and applications with competitive quality compared to
fixed access alternatives. Long Term Evolution (LTE) is the latest 3GPP stan-
dard, offering high peek data rates and low latency, currently under deployment
by many cellular network operators. The enhanced system capacity enables LTE
systems to be a valid alternative of fixed Internet access with the added value of
mobility. However, high data rate and low latency provided by LTE infrastruc-
ture are rarely enough to gain end-user satisfaction without higher layer services
being able to adapt to the environment and utilize resources in an optimal way.
Transmission Control Protocol (TCP) [6] is used by various Internet services
that require reliable end-to-end data transport for content delivery. Since TCP is
the dominant transport protocol of Internet based applications, its performance
plays a critical role in the user experience. Reliable and efficient data transfer
requires that TCP is able to dynamically adapt its rate to the available resources
and that it is capable to detect data loss in order to retransmit the missing data.
The rate adaptation is handled by TCP’s congestion control (CC) algorithm. The
main assumption of the TCP CC algorithm – that data loss is due to congestion
and not to bit errors – is valid in case of wired systems; however, limitations

R. Szabó and A. Vidács (Eds.): EUNICE 2012, LNCS 7479, pp. 263–274, 2012.

c IFIP International Federation for Information Processing 2012
264 V. Farkas, B. Héder, and Sz. Nováczki

introduced by the radio interface (large round-trip time (RTT), fluctuation of

the RTT, and transmission errors due to the imperfections of the wireless link)
make TCP non-optimal for transferring data over mobile networks [28].
The performance improvement of the TCP protocol over wireless links has
been a subject of intensive research in recent years. Proposals can be grouped
into two categories: end-to-end and network supported solutions. End-to-end so-
lutions require changes in the client and server TCP implementations, whereas
network supported solutions require extra functionality at the underlying net-
work. Important examples of end-to-end solutions are TCP Eifel [16] and TCP
Westwood [17]. The network supported solutions are commonly referred as Per-
formance Enhancing Proxies (PEP) [5]. The most prominent implementation of
the PEP concept is the split connection TCP proxy, which divides traversing
TCP connections into two independent connections, while being transparent to
the end-points. The TCP proxy can improve end-to-end performance of a TCP
connection is two ways. On one hand a proxy can separate the transmission
path along network boundaries with significantly different characteristics and
quarantine transmission problems in the affected loop. A proxy deployed at the
boundary of the wired and wireless part of a network can isolate the impacts of
the air interface problems (i.e., TCP congestion control problems) from the rest
of the system and handle these locally.
This paper evaluates the efficiency of the split connection TCP proxy concept
in LTE systems with proxy deployed in the System Architecture Evolution Gate-
way (SAE-GW). The performance is evaluated in the context of three mobile
applications, namely FTP based file transfer, web browsing and streaming video,
in terms of end-to-end throughput, mean download time and user satisfaction,
respectively. Numerical results are provided based on simulations.
The rest of the paper is organized as follows. In Section 2, an overview of the
related work is presented, which is followed by an introduction of the split TCP
proxy architecture in the LTE environment, along with the discussion of the
possible deployment options in Section 3. The simulation model and the results
of the simulations are discussed in Section 4. Finally, the paper is concluded in
Section 5.

2 Related Work

The concept and practical benefits of PEPs was in main focus of several research
projects during the last decade. Many research papers investigate the split con-
nection TCP proxy concept applied in case of satellite links that have high RTT
[8], ad-hoc networks with high bit error rate links [18], as well as in case of
transport infrastructures for reliable overlay multicast networks [15]. Lately sev-
eral studies investigate the potential benefits of using the split TCP concept in
wireless [27,13,22] and in particular in cellular networks [14]. The authors in [19]
give a comprehensive overview of split connection TCP proxies in Wideband
Code Division Multiple Access networks and show that this solution can signif-
icantly improve the throughput of file downloads, while in [21] the performance
 A Split Connection TCP Proxy in LTE Networks 265

(e.g., throughput, page download time) of Hypertext Transfer Protocol (HTTP)

proxy and TCP proxy solutions is compared. In [7] a TCP proxy with flow
aggregation capability is investigated applied in order to enhance the user ex-
perience in case of General Packet Radio Service data calls. A radio network
feedback controlled TCP proxy is studied in [20] that has as scope to improve
the end-to-end TCP performance in a 3G network. The radio quality feedback
(e.g., radio bandwidth changes, queue length in the Radio Network Controller)
is used to adapt the TCP window size in the proxy. A similar concept is studied
in [25] in order to mitigate the effects of bandwidth oscillation in High-Speed
Downlink Packet Access networks, while in [26] the same concept is empowered
by an analytical model and further simulation results.

3 TCP Proxy in LTE Networks

3.1 LTE Overview

The LTE system is structured into four high level domains (as depicted in Figure
1): User Equipment (UE), Evolved UMTS Radio Access Network (E-UTRAN),
Evolved Packet Core (EPC), and the Services domain [12]. The UE, E-UTRAN
and EPC form the Internet Protocol (IP) Connectivity Layer, with the main
function to provide IP based connectivity. E-UTRAN consists of the evolved
Node Bs (eNBs). The EPC consists of the Mobility Management Entity (MME),
which is the control plane functional element, and the SAE-GW (consisting of
the Serving-GW and Packet-GW), which is the user plane gateway to the E-
UTRAN. The Services domain consists of subsystems that provide the various
operator services or grants access to the external services such as those available
on the Internet. The UEs are connected to the eNBs via the air (Uu) interface.
eNBs are connected to the SAE-GW via the S1-U and to the MME via the S1-C
interface. The MME and the SAE gateway communicate with each other via
the S11 interface. User traffic is carried via the S1-U interface through General
Packet Radio Service Tunneling Protocol (GTP) tunnels, whereas control plane
traffic via the S1-C interface. The eNBs are interconnected via the X2 interface
that is used in order to transfer the handover related user and control plane
traffic.

3.2 The Split Connection TCP Proxy Concept

A split connection TCP proxy captures TCP session establishment attempts

and creates two independent TCP connections working in tandem instead of
one end-to-end session. The proxy transparently emulates the communication
end-points so that the presence of the proxy is seamless to them. The two legs
of the split connection are independent in means of flow and congestion control,
while they are using a shared buffer to exchange client data.
The main property of a split TCP connection is that RTT is reduced in
the split loops thus higher throughput can be achieved compared to the original
266 V. Farkas, B. Héder, and Sz. Nováczki

Fig. 1. The LTE network architecture

end-to-end throughput. As concluded in [3] the theoretical maximum end-to-end

throughput is the maximum throughput of the loop with higher RTT (i.e., the
one with smaller throughput). Reduced RTT enables higher maximum through-
put in the split loops, which can be achieved faster as TCP can increase its con-
gestion window size with an increased rate in all operation phases (slow start,
congestion avoidance and fast recovery). Other main benefit of a split TCP con-
nection is that effects of link failures on TCP performance are not propagated
between the loops, while the affected loop can recover faster. The latter property
can be highly beneficial in networks with wireless links where the main source
of packet drop is not network congestion but poor radio quality. The latter can
significantly deteriorate TCP performance as TCP assumes packet drops as in-
dication of network congestion.

3.3 Deployment Options

The benefits highlighted above can be exploited by deploying a TCP proxy
functionality in an LTE network, however the expected gains highly depend on
its location. The possible locations of the proxy functionality are the SAE-GW
(before S1 tunnel), the eNB (between the S1 tunnel and the Data Radio Bearer)
and finally the combination of the above two. Each solution has its drawbacks
and benefits, which are shortly outlined here.
The SAE-GW based solution (when the TCP proxy is located at the SAE-
GW) is the most straightforward option as the SAE-GW is the mobility anchor
point for user plane traffic. The RTT ratio of the split loops is more auspicious
as both split loops have smaller RTT than the original end-to-end RTT. As
shown in Figure 2 the smaller RTT enables shorter TCP connection setup and
data transfer times in slow-start mode. The drawback of this option is that air
interface problems are recovered slower compared to the eNB based solution.
An eNB based TCP proxy is the deployment option if air interface problems
need to be handled in a quick, low RTT radio loop and masked from the rest of
the network. However, the implementation is not as straightforward as in case of
the SAE-GW base solution as during handovers the TCP proxy context have to
be transfered between the eNBs. Moreover, as the RTT in the other loop (i.e.,
towards the CS) is almost the same as the original end-to-end RTT, the eNB
 A Split Connection TCP Proxy in LTE Networks 267

based TCP proxy cannot significantly increase end-to-end throughput in case of

good radio conditions on the air interface.
A combined SAE-GW/eNB solution is expected to deliver best results, as it
inherits the benefits of the base options while also complement each other. The
eNB proxy provides fast recovery on the air interface while the SAE-GW boosts
the end-to-end throughput. Note that the context transfers during handovers
still has to be managed in the eNB proxy.
Due to its simple implementation (i.e., no handovers had to be considered)
and its more general potential benefits (i.e., not bound to radio link failures) we
have selected the SAE-GW based option for our first investigations, the results
of which is presented in this paper. A SAE-GW side TCP proxy splits end-to-end
TCP connections into two corresponding TCP loops: one between the CS and
the SAE-GW and another between the SAE-GW and the UE. The TCP proxy
emulates the UE’s behavior towards the CS, while it faces as the CS towards
the UE. The presence of the proxy is transparent both to the UE and the CS.
Note, that the two TCP loops are independent from each other in means of
flow and congestion control. Moreover they might use different TCP parameters
or even variants. The loops are connected by a shared buffer in each direction.
Data received from the CS is stored in the downlink shared buffer as long as
it is forwarded to the UE, and a similar buffer is used for the uplink traffic. A
backpressure mechanism [15] is used to prevent buffer overflows.

Fig. 2. Duration of TCP connection setup (t1 an t2 ) and data transfer in slow start
phase (t3 and t4 ) with and without TCP proxy

4 Simulation Results

4.1 The Simulation Environment

The performance of the TCP proxy in LTE networks was investigated through
simulations in an event based simulation environment. In the simulator a reference
268 V. Farkas, B. Héder, and Sz. Nováczki

LTE network was created which simulates an urban environment with pedestrian
users and average radio conditions. All simulations were executed in this refer-
ence LTE network. The radio network layout of the reference LTE network was
designed according to the double ring network topology [1]; the 7 eNBs of the in-
ner ring were serving the traffic, the outer ring of eNBs were added only for in-
terference generation. In Figure 3 the topology of the reference LTE network is
depicted. The 7 eNBs were connected to the SAE-GW through IP routers, which
were connected via 100 Mbit/s Ethernet links with 10 ms media delay. The IP
routers used the Random Early Detection (RED) queue management technique
[10]. Each CS was connected through 100 Mbit/s Ethernet links with a fix delay
to the SAE-GW, this delay is referred as the mean Internet delay in the rest of the
paper. Each simulation case was executed with and without TCP proxy.

Fig. 3. The simulated topology

In the simulation environment the protocols of the S1-U and LTE-Uu inter-
face are implemented in detail. The air interface is based on 3GPP 25.814 [1]. A
single input single output channel with two times 10 MHz bandwidth is consid-
ered. Each eNB is simulated with a one cell one sector configuration, having an
omni-directional antenna. The mobility model is random way-point, inter-site
handover procedures are implemented. The channel quality is simulated sepa-
rately for each user considering distance loss, shadowing, multipath fading, and
penetration loss.
Three key Internet applications were used as benchmark applications in order
to evaluate the performance impacts of a TCP proxy in an LTE network: FTP
based file transfer (bulk data transfer), HTTP traffic (web browsing) and video
streaming. These three applications are implemented in detail in the simulation
environment. Each user had one of these three applications running during the
simulations. The TCP proxy was implemented in detail with two TCP variants:
New Reno and CUBIC. TCP New Reno [9] is also referred as standard TCP
because it is the most well-known TCP variant. TCP CUBIC [11] was developed
for networks with high bandwith-delay product and it is the current default TCP
variant of the Linux operating system so it is used by a large number of content
servers and Android mobile devices.
 A Split Connection TCP Proxy in LTE Networks 269

FTP users initiate a file download with unlimited file size. The measured per-
formance indicator for FTP traffic is the average throughput and the fairness
of the TCP connections as defined by the Jain’s Fairness Index [23]. The Jain’s
Fairness Index of connections with equal throughput is 1; the higher is the differ-
ence in the throughput of different connections, the lower is the Jain’s Fairness
Index.
The simulation model of web browsing was based on HTTP/1.1. Pipelining,
which has a large impact on the performance of mobile browsing was also part
of the model [4]. The browser client was modeled based on a typical mobile web
browser. Web browsing users were acting according to the common web surfing
behavior: they selected a random web page from a set of popular sites, down-
loaded them and proceeded to another random page after a simulated reading
time. Web pages were selected from the top 20 sites of Alexa’s top 500 global
list [2]. The structure of these web sites (i.e., the number of embedded objects
and the size of the HTTP request and response with the address of the server
for each object) was recorded in September, 2011. This state was used in the
simulations so that Domain Name System (DNS) queries, HTTP requests and
HTTP responses could be simulated authentically. For HTTP users, the mea-
sured performance indicator is the mean page download time.
Video streaming was modeled with the web browsing model extended with
the relevant features of the HTML5 (the fifth version of the HyperText Markup
Language) standard. The downloaded video had a 512 kbit/s average video bit-
rate and it was 90 s long. Videos of this size are not transferred with server side
rate control [24]. The performance measure in this case is the user satisfaction:
during the simulations users were considered satisfied if the buffering time of the
video was shorter than eight seconds and the video playback was not interrupted
more than three times or for more than two seconds altogether.

4.2 Numerical Results

First, the performance of downlink bulk data transfer with and without a TCP
proxy is evaluated. It is studied how the performance depends on the TCP vari-
ant, the mean Internet delay and the congestion level of the transport network.
FTP simulations were executed with both the New Reno and the CUBIC TCP
variant. For the mean Internet delay, three different values were set: 50, 75 and
125 ms; these represent the cases of communicating with a server with low,
medium and large delay. The different congestion levels of the transport net-
work were simulated with different number of users per cell. The not congested
case was modeled with 2 users per cell, the moderately congested case with 4
and 6 users per cell, and the congested case with 8 and 10 users per cell.
Figure 4 shows the per user average throughput and the Jain’s Fairness Index
in case of FTP data transfer with and without a TCP proxy when the mean
Internet delay was set to 50 ms. TCP proxy was able to increase the mean
throughput by more than 60% where there was no congestion, by 2-17% in case
of moderate congestion and by a few percents under heavy congestion. This
is due to the fact that the TCP proxy is not able to increase the throughput
270 V. Farkas, B. Héder, and Sz. Nováczki

of TCP connections if the transport network is congested, as in these cases the

load is at the maximum. The TCP variant had minimal impact on the achievable
throughput. The reason for this is that the bandwidth-delay product of the LTE
system is not in the range where the TCP CUBIC is able to achieve a higher
throughput than TCP New Reno. The Jain’s Fairness Index was above 0.96 in
all cases, which generally means a good fairness. However, the Fairness Index
was a bit worse with TCP proxy than without TCP proxy. The reason for this
is that during the simulations without TCP proxy a fix Internet delay is part
of the RTT of end-to-end TCP loops. On the other hand during simulations
with TCP proxy this fix delay is not part of the RTT of the SAE-GW–UE TCP
loop. Thus the variance of the RTT is smaller in the case with TCP proxy than
without TCP proxy.

4.5 1
New Reno w/o proxy
Mean throughput [Mbit/s]

4 Jain's Fairness Index

New Reno with proxy
3.5 0.98
CUBIC w/o proxy
3
CUBIC with proxy 0.96
2.5
2
0.94
1.5 New Reno w/o proxy
1 New Reno with proxy
0.92
0.5
CUBIC w/o proxy
CUBIC with proxy
0 0.9
2 4 6 8 10 2 4 6 8 10
Number of FTP connections per cell Number of FTP connections per cell

Fig. 4. Average throughput and Jain’s Fairness Index for FTP data transfer with and
without a TCP proxy when the mean Internet delay is 50 ms

4.5 1
New Reno w/o proxy
Mean throughput [Mbit/s]

4
Jain's Fairness Index

New Reno with proxy

3.5 0.98
CUBIC w/o proxy
3
CUBIC with proxy 0.96
2.5
2
0.94
1.5 New Reno w/o proxy
1 New Reno with proxy
0.92
0.5
CUBIC w/o proxy
CUBIC with proxy
0 0.9
2 4 6 8 10 2 4 6 8 10
Number of FTP connections per cell Number of FTP connections per cell

Fig. 5. Average throughput and Jain’s Fairness Index for FTP data transfer with and
without a TCP proxy when the mean Internet delay is 75 ms

Figure 5 and 6 show the same diagrams for the cases where the mean Internet
delay is 75 and 125 ms. These figures indicate a moderate gain in throughput
under light and moderate congestion. The smaller gain of TCP proxy in the
not congested case is because the larger RTT between the CS and the SAE-
GW decreases the achievable maximum throughput on the corresponding TCP
loop of TCP proxy. The larger increase in the moderately congested case is
 A Split Connection TCP Proxy in LTE Networks 271

4.5 1
New Reno w/o proxy
Mean throughput [Mbit/s]

Jain's Fairness Index

New Reno with proxy
3.5 0.98
CUBIC w/o proxy
3
CUBIC with proxy 0.96
2.5
2
0.94
1.5 New Reno w/o proxy
1 New Reno with proxy
0.92
0.5
CUBIC w/o proxy
CUBIC with proxy
0 0.9
2 4 6 8 10 2 4 6 8 10
Number of FTP connections per cell Number of FTP connections per cell

Fig. 6. Average throughput and Jain’s Fairness Index for FTP data transfer with and
without a TCP proxy when the mean Internet delay is 125 ms

possible because the larger Internet delay degrades the achievable throughput of
the reference cases, thus the transport network become fully utilized at a higher
number of users per cell.
The performance of web browsing with and without a TCP proxy is presented
for the New Reno TCP variant. All HTTP simulations were repeated with the
CUBIC TCP variant as well but the performance difference between the two
variants was minimal, so those results are not shown. The investigated mean
Internet delay values were the same as for the FTP simulations. The congestion
level of the network was changed by adding background FTP downloads; in the
uncongested case, there were no background FTP downloads while 2 to 4 and 6
to 8 FTP downloads per cell simulated the moderately congested and congested
cases. Every HTTP simulation case was executed with 5 different random seeds.
Figure 7 (a) shows the mean download times when the mean Internet delay
was 50 ms. It is clear that the TCP proxy decreases the average download time
in all the investigated cases. As expected, the mean download time increases
when the transport gets more congested but there is a benefit of a TCP proxy
in those cases as well.

ϳ ϳ ϳ
Mean download time [s]

Mean download time [s]

Mean download time [s]

ϲ ϲ ϲ
ϱ ϱ ϱ
ϰ ϰ ϰ
ϯ ϯ ϯ
Ϯ Ϯ Ϯ
ϭ ϭ ϭ
w/o proxy with proxy w/o proxy with proxy w/o proxy with proxy
Ϭ Ϭ Ϭ
Ϭ Ϯ ϰ ϲ ϴ Ϭ Ϯ ϰ ϲ ϴ Ϭ Ϯ ϰ ϲ ϴ
Background load Background load Background load
(# of FTP downloads per cell) (# of FTP downloads per cell) (# of FTP downloads per cell)

         

Fig. 7. Mean download time of web pages with and without a TCP proxy when the
mean Internet delay is (a) 50 ms, (b) 75 ms and (c) 125 ms
272 V. Farkas, B. Héder, and Sz. Nováczki

ϭϬϬ ϭϬϬ ϭϬϬ

User satisfact. ratio [%]

User satisfact. ratio [%]
User satisfact. ratio [%]

ϵϱ ϵϱ ϵϱ

ϵϬ ϵϬ ϵϬ
ϴϱ ϴϱ ϴϱ
ϴϬ ϴϬ ϴϬ
ϳϱ ϳϱ ϳϱ
w/o proxy with proxy w/o proxy with proxy w/o proxy with proxy
ϳϬ ϳϬ ϳϬ
Ϭ Ϯ ϰ ϲ ϴ Ϭ Ϯ ϰ ϲ ϴ Ϭ Ϯ ϰϲ ϴ
Background load Background load Background load
(# of FTP downloads per cell) (# of FTP downloads per cell) (# of FTP downloads per cell)

         

Fig. 8. The ratio of satisfactory and all video playbacks with and without a TCP proxy
when the mean Internet delay is (a) 50 ms, (b) 75 ms and (c) 125 ms

Figure 7 (b) and (c) shows the same diagrams for higher Internet delays. As ex-
pected, the average download time increases with the mean Internet delay.
However, TCP proxy can decrease the download time by the same amount (which
is obviously a smaller ratio compared to the value of the download time), there-
fore the performance gain of TCP proxy decreases in case of web browsing as well
when the Internet delay increases.
The performance of video streaming with and without TCP proxy is pre-
sented on Figure 8. The investigated conditions were the same as for the HTTP
simulations. Each simulation case contained 228 video downloads and they were
executed with 4 different random seeds, so from altogether 896 video downloads
a user satisfaction ratio was calculated. The figures indicate that TCP proxy
improves the user satisfaction ratio of video streaming in all the investigated
conditions. The improvement is more profound when the Internet delay or the
congestion level of the network is higher.

5 Conclusion

In this paper, it has been presented how the split connection TCP proxy con-
cept can be used in LTE networks. Furthermore, it has been evaluated how the
SAE-GW based TCP proxy affects the performance of bulk data transfer, web
browsing and video download. Numerical results show that the TCP proxy can
considerably increase the performance of web browsing and video download in
all investigated conditions. TCP proxy is also able to significantly increase the
performance of bulk data transfer if the transport network is not congested;
however, there is no improvement when the transport is already fully utilized
and the TCP proxy can even decrease the fairness in this case.
A possible extension of the work presented in this paper could be the evalua-
tion of the eNB based TCP proxy solution. The study of the combined solution
(two TCP proxies, one in the eNB and another in the SAE-GW) could also be an
interesting follow-up topic. An additional research topic could be the fine tuning
 A Split Connection TCP Proxy in LTE Networks 273

of the TCP parameters of the TCP proxy (e.g. initial window size) in a static
or an adaptive way based on current network conditions.

References

1. 3GPP TS 25.814: Physical layer aspect for evolved Universal Terrestrial Radio
Access, UTRA (2006)
2. Alexa: The top 500 sites on the web, http://www.alexa.com/topsites/
3. Baccelli, F., Carofiglio, G., Foss, S.: Proxy caching in split tcp: Dynamics, stabil-
ity and tail asymptotics. In: The 27th Conference on Computer Communications,
INFOCOM 2008, pp. 131–135. IEEE (April 2008)
4. Blaze: HTTP Pipelining – Big in Mobile,
http://www.blaze.io/mobile/http-pipelining-big-in-mobile/
5. Border, J., Kojo, M., Griner, J., Montenegro, G., Shelby, Z.: RFC 3135: Perfor-
mance Enhancing Proxies Intended to Mitigate Link-Related Degradations. Inter-
net Engineering Task Force (June 2001)
6. Braden, R.: RFC 1122: Requirements for Internet Hosts - Communication Layers.
Internet Engineering Task Force (October 1989)
7. Chakravorty, R., Katti, S., Pratt, I., Crowcroft, J.: Using tcp flow-aggregation to
enhance data experience of cellular wireless users. IEEE Journal on Selected Areas
in Communications 23(6), 1190–1204 (2005)
8. Dubois, E., Fasson, J., Donny, C., Chaput, E.: Enhancing tcp based communi-
cations in mobile satellite scenarios: Tcp peps issues and solutions. In: 2010 5th
Advanced Satellite Multimedia Systems Conference (ASMA) and the 11th Signal
Processing for Space Communications Workshop (SPSC), pp. 476–483 (September
2010)
9. Floyd, S., Henderson, T.: RFC 2582: The NewReno Modification to TCP’s Fast
Recovery Algorithm. Internet Engineering Task Force (April 1999)
10. Floyd, S., Jacobson, V.: Random early detection gateways for congestion avoidance.
IEEE/ACM Trans. Netw. 1(4), 397–413 (1993)
11. Ha, S., Rhee, I., Xu, L.: CUBIC: a new TCP-friendly high-speed TCP variant.
SIGOPS Oper. Syst. Rev. 42(5), 64–74 (2008)
12. Harri Holma, A.T.: LTE for UMTS: OFDMA and SC-FDMA Based Radio Access.
Wiley (2009)
13. Hoque, M., Siekkinen, M., Nurminen, J.: On the energy efficiency of proxy-based
traffic shaping for mobile audio streaming. In: 2011 IEEE Consumer Communica-
tions and Networking Conference (CCNC), pp. 891–895 (January 2011)
14. Ivanovich, M., Bickerdike, P., Li, J.: On tcp performance enhancing proxies in a
wireless environment. IEEE Communications Magazine 46(9), 76–83 (2008)
15. Kwon, G.I., Byers, J.: Roma: reliable overlay multicast with loosely coupled tcp
connections. In: Twenty-third AnnualJoint Conference of the IEEE Computer and
Communications Societies. INFOCOM 2004, vol. 1, vol. (xxxv+2866), p. 4 (March
2004)
16. Ludwig, R., Katz, R.H.: The Eifel algorithm: making TCP robust against spurious
retransmissions. SIGCOMM Comput. Commun. Rev. 30(1), 30–36 (2000)
17. Mascolo, S., Casetti, C., Gerla, M., Sanadidi, M.Y., Wang, R.: TCP westwood:
Bandwidth estimation for enhanced transport over wireless links. In: MOBICOM,
pp. 287–297 (2001)
274 V. Farkas, B. Héder, and Sz. Nováczki

18. Matsumoto, M., Yamamoto, M.: Performance evaluation of flow fairness in proxy-
based tcp for ad hoc networks. In: 2011 IEEE Pacific Rim Conference on Com-
munications, Computers and Signal Processing (PacRim), pp. 780–785 (August
2011)
19. Meyer, M., Sachs, J., Holzke, M.: Performance evaluation of a tcp proxy in wcdma
networks. IEEE Wireless Communications 10(5), 70–79 (2003)
20. Moller, N., Molero, I., Johansson, K., Petersson, J., Skog, R., Arvidsson, A.: Using
radio network feedback to improve tcp performance over cellular networks. In:
44th IEEE Conference on Decision and Control, 2005 and 2005 European Control
Conference, CDC-ECC 2005, pp. 7434–7439 (December 2005)
21. Necker, M.C., Scharf, M., Weber, A.: Performance of Different Proxy Con-
cepts in UMTS Networks. In: Kotsis, G., Spaniol, O. (eds.) Mobile and
Wireless System. LNCS, vol. 3427, pp. 36–51. Springer, Heidelberg (2005),
http://dx.doi.org/10.1007/978-3-540-31963-4_4
22. Pu, W., Zou, Z., Chen, C.W.: New tcp video streaming proxy design for last-
hop wireless networks. In: 2011 18th IEEE International Conference on Image
Processing (ICIP), pp. 2225–2228 (September 2011)
23. Jain, R., Chiu, D., Hawe, W.: A Quantitative Measure Of Fairness And Discrimina-
tion For Resource Allocation In Shared Computer Systems. DEC Research Report
TR-301 (September 1984)
24. Rao, A., Legout, A., Lim, Y.S., Towsley, D., Barakat, C., Dabbous, W.: Network
characteristics of video streaming traffic. In: Proceedings of the Seventh Conference
on Emerging Networking EXperiments and Technologies, CoNEXT 2011, pp. 25:1–
25:12. ACM, New York (2011), http://doi.acm.org/10.1145/2079296.2079321
25. Ren, F., Huang, X., Liu, F., Lin, C.: Improving tcp throughput over hsdpa net-
works. IEEE Transactions on Wireless Communications 7(6), 1993–1998 (2008)
26. Ren, F., Lin, C.: Modeling and improving tcp performance over cellular link with
variable bandwidth. IEEE Transactions on Mobile Computing 10(8), 1057–1070
(2011)
27. Shikama, T.: Mitigation of bursty packets by a tcp proxy improving tcp perfor-
mance in a wired and wireless network. In: 2010 IEEE GLOBECOM Workshops
(GC Wkshps), pp. 425–429 (December 2010)
28. Xylomenos, G., Polyzos, C., Mahonen, P., Saaranen, M.: TCP performance issues
over wireless links. IEEE Communications Magazine 39(4), 52–58 (2001)