IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.

6, JUNE 2008 1993

Improving TCP Throughput over HSDPA Networks

Fengyuan Ren, Member, IEEE, Xiaomeng Huang, Feng Liu, and Chuang Lin, Senior Member, IEEE

Abstract—The various link adaptation techniques employed Recently, researchers have begun to investigate the impact
by High Speed Downlink Packet Access (HSDPA) in the third of the wireless link bandwidth variation on the TCP per-
generation (3G) networks augment the bandwidth oscillation, formance [3][4][5][6]. 3G networks are required to support
which is identified as one of the most important factors resulting
in the throughput deterioration of Transmission Control Protocol dynamic resource sharing among concurrent data users. If
(TCP). In this paper, we firstly explain why the bandwidth multiple users wish to transfer large numbers of data simulta-
oscillation degrades the TCP performance through a special neously, the scheduler may have to repeatedly allocate and de-
simulation experiment. Subsequently, a split connection Win- allocate resources for each user. Intuitively, the channel-state
dow Adaptation TCP Proxy is proposed to improve the TCP based scheduling mechanisms improve the overall throughput,
throughput over HSDPA networks. In this solution, the built-in
attributes of HSDAP system are sufficiently utilized. In order while they also result in bandwidth oscillations, which were
to effectively use the precious cellular link resources, the length identified as one of the most important factors in deteriorat-
of the queue connected with it is intentionally kept around the ing throughput [3][6].It is an open question to what extent
reference value through adjusting the sending window size of transport protocols can be designed to optimize performance
TCP proxy based on the dynamic values of varying bandwidth. in the presence of bandwidth oscillation, and to what extent
A discrete-time stochastic state space model is formulated to
analyze the system stability. The validity of enhanced scheme wireless link-level mechanisms for bandwidth variation can be
is verified through simulation experiments. The performance of designed to take into account the transport protocols that will
TCP proxy is compared with the standard TCP protocol. The run over these links [7].
numerical results show that our TCP proxy is able to keep In this paper, we will restrict ourselves to Universal Mobile
the cellular link utilization over 90%, and to improve TCP Telecommunication System(UMTS) based 3G networks to
throughput by 100% under most conditions.
investigate this open performance optimization issue. Based
Index Terms—TCP performance, bandwidth oscillation, proxy, on the split-connection approach [8], we propose a new
stochastic control, HSDPA networks. TCP enhancement solution called Window Adaptation TCP
Proxy to maximize TCP throughput over HSDPA in UTMS
I. I NTRODUCTION networks. Our TCP proxy sufficiently utilizes the inherent
attribute of HSDAP system to obtain the dynamic bandwidth

T HE wireless data services are anticipated to have a rapid

growth over the next years, and will likely become the
dominating source of traffic load in 3G wireless cellular
value, and employs a dynamic window adjusting scheme to
adapt to the variable bandwidth.
The remainder of the paper is organized as follows. The
networks, which are increasingly being deployed throughout simplified HSDPA operation principle and its technical fea-
the world. Therefore, an efficient provision for the expected tures are introduced, and the related works are discussed
diverse data services is considered a key factor for the success in Section 2. In Section 3, why the bandwidth oscillation
of 3G networks. Since the vast majority of data applications degrades the TCP performance is analyzed and explained
rely on TCP in the Internet, particular attention must be paid through a special simulation experiment. In Section 4, the
to the performance of TCP over 3G wireless networks. architecture of window adaptation TCP proxy is described
In fact, the performance of TCP over wireless networks and its advantages are discussed. Subsequently, the control
has been extensively studied in last decade since the experi- algorithm is designed to dynamically regulate the sending
ment results show that TCP suffers the significant throughput window size of TCP proxy, and then a discrete-time stochastic
degradation when there is a wireless link in the end-to-end state space model is formulated to analyze the system stability.
connection [1]. It has been found that wireless link losses In Section 5, the validity of our TCP Proxy is verified and
have an adverse impact on performance as TCP is unable to its performance is compared with the standard TCP through
distinguish corruption loss from congestion loss, so any kind simulation experiments. Finally, the conclusions are drawn in
of losses are perceived as congestion, which results in source Section 6.
throttling and very low throughout. These findings have been
one of the main motivations for the use of extensive local II. BACKGROUND AND R ELATED W ORKS
retransmission mechanisms in 3G networks[2].
A. HSDPA
Manuscript received March 12, 2007; revised September 19, 2007; accepted Firstly, we provide a general overview of the HSDPA tech-
November 9, 2007. The associate editor coordinating the review of this paper nology since it may be helpful to achieve a full comprehension
and approving it for publication was W. Zhuang.
F. Ren, X. Huang, and C. Lin are with the Computer Science and of the investigation in this paper. HSDPA is a new technology
Technology Department, Tsinghua University, Beijing, P.R. China (e-mail: standardized by 3GPP (3G Partnership Project) in Release
[email protected], [email protected]). 5 to boost the support for packet switched services as an
F. Liu is with the School of Electronics and Information Engineering,
Beihang University, Beijing, P.R. China. evolution of UMTS radio interface. The main idea behind the
Digital Object Identifier 10.1109/TWC.2008.061007. HSDPA is to share one single downlink data channel among
1536-1276/08$25.00 
c 2008 IEEE
1994 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

UE1 HS-SCCH UE Node-B/RNC Server

Demodulation & HARQ Info, UE ID.
bw, 10ms 10Mbps, 5ms
HS-DSCH Downlink User 1
HS- PDSCH1
cellular link wire link
UE2 User 2
Fig. 2. Simulation topology.
...

...

...
HS-PDSCH15 User 3

UE3 HS-DPCCH are finally attributed to the spurious retransmission. However,

CQI CQI ACK CQI Node -B it has been mentioned that no performance benefit has been
Uplink
Ass. DCH obtained with modifications in the TCP retransmission timer in
Power Control Commands
[9]. M.C. Chan and R. Ramjee believed that TCP performance
Fig. 1. HSDPA operation principle.
degradation caused by bandwidth and delay variation is due to
the difficulty in estimation of bandwidth-delay product(BDP)
of the end-to-end path at the TCP source [3]. They developed
all the users, along with many modifications that make it two improving schemes successively, i.e. ACK Regulator [3]
especially effective for packet-based communication. HSDPA and Window Regulator [4]. The ACK Regulator runs in an
utilizes Adaptive Modulation and Coding (AMC) to substitute intermediate bottleneck network element and monitors the
fast power control and variable spreading factor in traditional arrival rate of packets, and controls the release of ACKs to
3G technology. To facilitate the high peak data rates, HSDPA the TCP source so as to prevent buffer overflow. The main
introduces the variable-rate turbo encoding and employs the advantage of the ACK Regulator is that no modification is
multi-code transmission. The peak rate varies from 120 kbps needed at end systems. However, it does not address the
up to 10.8 Mbps (in practice up to 2 Mbps) under different performance of short-lived flows. The Window Regulator is
parameter configurations. very similar with the ACK Regulator. It manages to maintain
The simplified HSDPA operation principle is illustrated the queue length at the bottleneck link within the desired
in Fig.1. Depending on packet prioritization and resource range through dynamically computing the adequate receive
availability, Node-B schedules user data on the High Speed window value in the ACKs sent from the receiver back to the
Downlink Shared Channel (HS-DSCH), where a large amount sender. However, the algorithms in [3] and [4] have a common
of power and code resources are assigned to a single user at a features, namely they infer the bandwidth variation through
certain TTI(Transmit Time Interval) in time and/or code mul- observing the evolution of queue length. This approach is ef-
tiplex manner. Prior to sending data on the HS-DSCH, Node- fective for slow and moderate variation in bandwidth, but does
B sends a detailed demodulation message to the active users not work for rapid and drastic bandwidth oscillations since the
via the High-Speed Shared Control Channel (HS-SCCH). buffer may become empty or full in this situation. Naturally,
The uplink High Speed Dedicated Physical Control Channel it is hard to calculate the appropriate receive window value
(HS-DPCCH) carries the necessary control information, such in ACKs or the release rate of ACKs using the heuristic
as ARQ acknowledgements and Channel Quality Indicators algorithms with the incomplete bandwidth information. In
(CQIs). CQI is an indicator of AMC and multi-code number addition, the determinate analysis techniques used in these
currently supported by the user equipment (UE), namely the work can not provide comprehensive evaluation on system
supported data rate under the current channel conditions. The performance since the bandwidth variation is random and
feedback cycle of the HS-DPCCH CQI can be set as a network indeterminate.
parameter in predefined steps, such as 2ms and 20ms etc.
On the viewpoint of optimizing TCP performance, there III. W HY BANDWIDTH O SCILLATION D EGRADES TCP
are two obvious features in HSDPA technology. First, the P ERFORMANCE
various link adaptation techniques employed by HSDPA give
rise to the bandwidth variation. Otherwise, the fact that the In order to focus on the effects of bandwidth oscilla-
bandwidth of the downlink data channel is shared among all tion on TCP performance, we use a rather simple network
users also augments high bandwidth variation. The available topology shown in Fig.2 to conduct an experiment on the
bandwidth for a connection is affected by the number of ns2 simulation platform since it may neglect the effects
users entering and leaving the cell since they all compete caused by other possible factors. The system consisted of
for the resources. Second, some wireless link-level enhanced an integrated NodeB/RNC, a server connected to the RNC
mechanisms, such as AMC and scheduling schemes based on by a 10 Mbps wired link, and a mobile device connected
CQI, make NodeB accurately know the bandwidth assigned to to the NodeB by a cellular link with 96 kbps uplink and
each user. Although it is not purposely designed for optimizing varying bandwidth downlink, whose capacity varies according
TPC performance, we will adequately exploit this potential in to the rectangular pattern shown in Fig.3(a), i.e. the interval
our TCP proxy. is 10s, and the peak bandwidth and the valley bandwidth
are 384 kbps and 38.4 kbps respectively. The TCP source at
the server using newReno transfers a large file to the mobile
B. Related work device. The maximum transmission unit (MTU) is 1500bytes
In [6], F. Khafizov and M. Yavuz conducted experiments of and the buffer at the RNC is 20 packets. We monitor the
TCP over IS-2000 networks and concluded that bandwidth evolutions of the congestion window at TCP source and the
oscillation severely degrades TCP throughput. The reasons queue length at RNC, and trace TCP sequences (including
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008 1995

Fig. 3. Simulation Experiment plot (a) shows the bandwidth variation, Fig. 4. TCP sequence trace.
plot (b) shows the evolution of congestion window, and plot (c) shows the
evolution of queue length in the intermediate node.

bandwidth for each UE based on channel state and maintains a

dedicated queue for each flow. We will sufficiently utilize the
sending packets, receiving ACKs, and dropping packets). inherent attribute to improving TCP performance over HSDAP
The results are presented in Fig.3(b), Fig.3(c) and Fig.4 networks with bandwidth oscillation.
respectively.
 150 Theoretically, the average throughput should be
equal to 0 bw(t)dt/150 = 222.7kbps, but the actual average IV. O PTIMIZING TCP P ERFORMANCE IN HSDPA
throughput is 190.4 kbps, and the link utilization is about 85%.
The split connection scheme proposed in [8] has been
Observing Fig.3 and Fig.4, we can find some hints that the
widely used in commercial cellular networks [11]. In this
bandwidth oscillation deteriorates TCP performance. At time
work, combining with the architecture of UTMS network and
t=50 and 90, the link bandwidth decreases from 384 kbps to
the inherent attribute of HSDPA system, we propose a split
38.4 kbps, the buffer overflows, and the congestion window
connection Window Adaptation TCP proxy to optimize TCP
is shrunk, which indicates that the Fast Retransmission mech-
performance.
anism is triggered. Subsequently, the congestion window is
reset to 1 but without any dropping packets, which means that
the spurious timeouts occurs. This phenomenon agrees with A. Architecture
the explanation presented in [6]. However, not all bandwidth A simplified architecture of UMTS network is depicted in
attenuations certainly result in the spurious timeouts, such as at Fig.5, only the elements that play an important role on the
time t=30 and 70, since the queue length is relatively small at TCP connection are presented. The UMTS functionality is
that time, the variation of queuing delay caused by the abrupt divided into three main domains: UE, UMTS Terrestrial Radio
decrease bandwidth is not large enough to lead the spurious Access Network (UTRAN) and Core Network (CN). The
timeouts. It is noticeable that the bandwidth is decreasing at Core Network provides switching and routing for user traffic,
time t = 30, but the congestion window is increasing, finally and comprises two basic nodes: Serving GPRS Support Node
the buffer overflows so that the window size is halved. Next, (SGSN) and Gateway GPRS Support Node (GGSN). The latter
the bandwidth jumps, but the congestion window is too small provides access to external networks, such as the Internet.
to fill the link, moreover, it climbs slowly, so that the scarce UTRAN consists of Node-B and RNC, and is responsible for
wireless link resource is wasted. Therefore, it is incomplete if providing the UEs with access to the CN, as well as managing
TCP throughput deterioration caused by bandwidth oscillation the radio resources. Node-B maintains the per-user queue and
is merely attributed to the spurious timeouts. We believe that performs the link adaptation operation, and then schedules
the essential reason should be that the congestion window transmission on the wireless channel based on the channel
does not timely catch up the varying bandwidth. If they could state. Our split connection TCP proxy is located between
harmoniously match, the link would be fully utilized, and there GGSN and the external networks. For the convenience of
should not be so many packets to accumulate in the queue, discussion, the basic structure of our window adaptation TCP
then the spurious timeouts would be naturally avoided. The proxy is illustrated in Fig.6. One TCP connection between the
necessary condition that TCP sending window timely keeps UE and the server is split into two ones: one between the UE
up with the bandwidth changes is that the accurate bandwidth and the proxy and another between the proxy and the server.
information available. It is hard to estimate the available The connection between the server and proxy can employ
bandwidth in Internet. Fortunately, it is readily obtained in any standard TCP protocols, while the transport protocol on
HSDPA network because Node-B is responsible for allocate the connection between the proxy and the UE needs to be
1996 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

bw UDP(bw, q)
U Node-B Controller

Server
UE1

UE2

Cell
UE5
IP Networks E
q
UE3
UE4 Node-B RNC SGSN GGSN TCP Proxy
Modified TCP TCP proxy Standard TCP
UTRAN Core Networks

Fig. 6. Basic structure of window adaptation TCP proxy.

Fig. 5. Simplified UTMS network architecture.

variables, whose mean and variance are equal to 0 and 1

customized in order to adapt to the dynamic changes of respectively. Both αi (i = 1 . . . p) and βj (j = 1 . . . q) are
the wireless link bandwidth. For the TCP connection on the constant. ARMA models can adequately model any stationary
wireless link, the proxy is a sender. All the schemes and time series, at the same time, they are analytically tractable
algorithms related with congestion window adjustment in the [12]. Although Node-B explicitly feeds back the available
standard TCP will be disabled in the modified TCP protocol. bandwidth and the queue length to the TCP proxy, the proxy
The sending window size is dynamically regulated depending can only obtain the noisy estimates of these two random
on the available wireless bandwidth and the queue length in variables due to the interferences introduced by all kinds of
the Node-B. They are periodically conveyed to the TCP proxy sources, such as transfer delay and loss of signaling packet.
through the UDP connection established between Node-B and Let b̂(n) denote the estimate for the available bandwidth, and
the proxy. If the queue length in Node-B could be kept around q̂(n) denote the estimate for the queue length. Assume that
the desired small value all along, the perfect performance these estimates can be expressed as:
would be reached since both buffer overflow and empty queue
can be avoided so that the cellular link will be fully utilized. b̂(n) = b(n) + γ(n + 1) (3)
To achieve this target, we need to design a stable and effective q̂(n) = q(n) + η(n + 1) (4)
algorithm to control the sending window size, which will be
discussed in detail in the next section. Since the flow control where γ(n + 1) and η(n + 1) are Gaussian white noise, thus
algorithms in the standard TCP, including slow start, fast the TCP proxy can use these estimates to update its sending
retransmission, fast recovery and timeout retransmission, are window size.
disabled in our modified TCP version, most possible factors Substituting (3) and (4) into (1), we have:
resulting in performance degradation over wireless cellular
q(n + 1) = q̂(n) + w(n) − b̂(n) + π(n + 1) (5)
links, such as superior timeouts and small initial window size
etc., are naturally eliminated. However, the recovery strategy where π(n + 1) = η(n + 1) − γ(n + 1), it is another Gaussian
in the standard TCP will be retained, namely 3 duplicate ACKs white noise. In order to solve the optimal window adjusting
trigger the retransmission of the corresponding packet, but it algorithm, we define the cost function J as:
does not affect the sending window adjustment.
J = E[q(n + 1) − q0 ]2 (6)
B. Algorithms design where q0 is the desired queue length. Through minimizing the
To design a stable and effective flow control algorithm cost function J, we can get the adjusting algorithm of sending
running on the TCP proxy, we firstly build a discrete-time window size on the TCP proxy as follows:
stochastic model to describe the dynamic behavior of the w(n) = b̂(n) − q̂(n) + q0 (7)
system. The time is divided into steps, which correspond to the
intervals at which the feedback information is sent on the UDP For the TCP connection on the wired link, the TCP proxy
connection. At the end of each step, the TCP proxy updates its is a receiver. Without any control, the server will utilize the
sending window size using the feedback information, i.e. the abundant bandwidth to rapidly send packets, and there will
available bandwidth and the queue length. Let w(n) denote be a lot of packets in the buffer on the TCP proxy, so that
the sending window size during step n. Let b(n) and q(n) the end-to-end delay increases and some packets in real-time
denote the bandwidth of radio channel and the instantaneous interactive applications may become stale. In order to avoid
queue length in Node-B at the end of the nth step respectively. this possible phenomenon, we employ the advertised window
Defining the duration of each step as a unit time, we have: inserted into the corresponding field of ACKs to limit the
sending rate of the server. The objective is still to keep the
q(n + 1) = q(n) + w(n) − b(n) (1) queue length around the reference value. Here, we assume the
time-varying sending window of the TCP proxy as a stochastic
Since the available bandwidth b(n) may change over time process. Following the same way, we can obtain the control
randomly, we model it as an ARMA(p, q) stochastic process. law to regulate the advertised window size:
p
 q

b(n) = b + ϕ(n) = b + αi ϕ(n − i) + βj φ(n − j) (2) aw(n) = w(n) − q2 (n) + q20 (8)
i=1 j=1
where aw(n) denotes the advertised window size, q2 (n) is the
where b denotes the nominal value, φ(n) is independent instantaneous queue length on the TCP proxy, and q20 is its
and identically distribution(i.i.d.) standard Gaussian random reference value.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008 1997

⎛ ⎞
C. Stability analysis α1 α2 · · · αp−1 αp
  ⎜ 1 0 ··· 0 0 ⎟
The control law (7) and (8) will be effective as long as 0 −1 ⎜ ⎟
B= C=⎜ .. .. .. .. ⎟
the system is stable. In this section, we will proof a theorem 1 1 ⎝ . . . . ⎠
relating to the
system stability. 0 0 ··· 1 0
p Matrix Ahas p + 2 eigenvalues. The two eigenvalues should
Theorem If i=1 |αi | < 1 and E[q(0)] = q0 then the
following system is neutrally stable. be determined by matrix B, and the other p eigenvalues are
fixed by the matrix C. The eigenvalues of the matrix C are the
q(n + 1) = q(n) + w(n) − b(n) (9) roots of a complicated polynomial of degree p, and it is hard
p q to obtain the explicit solutions. However, using Gerschgorin’s
 
b(n) = b + αi ϕ(n − i) + βj φ(n − j) (10) theorem[13], we can find that the modulus of any root of this
i=1 j=1 polynomial satisfies the following inequality:
p

w(n) = b(n) − q(n) + q0 + π(n + 1) (11)
|λ − α1 | ≤ αi (17)
i=2
p
Proof For the convenience of analysis, the system can be Thus i=1 |αi | < 1 is a sufficient condition that the system
rewritten⎛as the state space⎞model[10].
⎛ Let ⎞ is stable. If this condition holds, as n  1 , the corresponding
w(n) − b π(n + 1) terms in (16) approaches zero, the value of E[X(n)] is only
⎜ q(n) − q0 ⎟ ⎜ 0 ⎟
⎜ ⎟ ⎜ q ⎟ determined by the eigenvalues of B.
⎜ ϕ(n) ⎟ ⎜ β φ(n − j) ⎟ 2
X(n)=⎜ ⎟ Y(n)= ⎜ ⎟ √ det(B − λI) = λ − λ + 1 = 0, we obtain λ1,2 =
From
j=0 j
⎜ .. ⎟ ⎜ .. ⎟
⎝ . ⎠ ⎝ . ⎠ (1 ± 3i)/2. Since |λ1,2 | = 1 , the system is neutrally stable
ϕ(n − p + 1) according to the conclusion in [13].
⎛ ⎞ 0
0 −1 1 0 ··· 0 0
⎜ 1 1 −1 0 · · · 0 0 ⎟ V. S IMULATION R ESULTS
⎜ ⎟
⎜ 0 0 α1 α2 · · · αp−1 αp ⎟
A=⎜ ⎟ We have implemented the Window Adaptation TCP Proxy
⎜ .. .. .. .. .. .. ⎟
⎝ . . . . . . ⎠ on ns2.29 simulator and conducted the simulation experiments
0 0 0 0 ··· 1 0 to validate its performance and made comparison with the
then it is easy to verify that the system can be rewritten as: standard TCP. Our TCP Proxy is inserted between Node-B and
the server depicted in Fig.2. We use a random variable, whose
X(n + 1) = AX(n) + Y(n) (12) peak value is 2Mbps and the varying period is T = 200 ms,
to describe the actual bandwidth oscillation shown in Fig.7(a).
The general solution of equation (12) is:
Without special statement, the parameters in foregoing simu-
n−1
 lation are kept unchangeable. We assume the updating step
X(n + 1) = An X(0) + An−i−1 Y(i) (13) as the unit time in the previous analysis and design. If the
i=1 sampling interval is τ , the corresponding control law will
Taking expected values on both sides of equation (13), we become:
yield: w(n) = b̂(n)τ − q̂(n) + q0 (18)
n−1
 Let τ = 10ms, and fix q0 and q20 at 15 packets and 50
E[X(n + 1)] = An E[X(0)] + An−i−1 E[Y(i)] (14) packets respectively. The whole simulation lasts 150s. The
i=1
related parameters are traced and presented in Fig.7 together.
Since the random variables π(n) and φ(n) are zero-mean, Fig.7(b) and Fig7 (d) describes the sending window size of
equation (14) becomes: the proxy and the queue length in Node-B controlled by our
Window Adaptation TCP Proxy. When the server employs
E[X(n + 1)] = An E[X(0)] (15) TCP NewReno without our TCP Proxy, the evolutions of the
If A can be diagonalized, i.e. A = SΛS−1 ,then congestion window and the queue length in Node-B are shown
in Fig.7 (c) and Fig.7(e) respectively. The congestion window
E[X(n+1)] = SΛS−1 E[X(0)] = k1 λn1 f1 +· · ·+kp+2 λnp+2 fp+2 of TCP newReno displays the regular saw-tooth shape, which
(16) does not timely catch up the dynamic change of the bandwidth
where [λ1 , · · · , λp+2 ] is eigenvalues of the matrix A, and and results in queue oscillation with large amplitude. Plenty of
the coefficients f = [f1 , · · · , fp+2 ]T that match the initial packets are dropped due to buffer overflow, and the end-to-end
condition E[X(0)] are f = S−1 E[X(0)]. goodput is deteriorated. On the other hand, the empty queue
The general solution (16) of equation (12) indicates that occurs frequently, the wireless link is in underutilization. On
the stability of the system (9) (11) can be analyzed using the the contrary, our TCP proxy dynamically adjusts the sending
eigenvalues of the matrix A , Observing the structure of A, window based on the queue length on the Node-B and the
some characteristics are beneficial to analyzing its eigenvalues. cellular link bandwidth. The queue length is always kept
We can decompose the matrix A into blocks matrices. The around the constant value. Both packet dropping and empty
upper left 2 × 2 block matrix of A is denoted by B, and the queue seldom appear. The end-to-end performance, including
lower right p × p block matrix of A is denoted by C, namely: goodput and delay jitter, is improved. The upper limit of
1998 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

enhancement proxy to optimize TCP throughput over HSDPA

network with bandwidth oscillation. The sending window size
of our TCP proxy is regulated based on the dynamic values
of varying bandwidth. Buffer overflow and empty queue are
effectively prevented through holding the length of the queue
connected with the cellular link around a fixed value, so
that the bottleneck wireless link is fully utilized and TCP
throughput is improved. Since TCP proxy adopts a completely
new flow control mechanism, and the traditional algorithms in
the standard TCP are disabled, many possible factors resulting
in performance degradation over wireless cellular links also
naturally disappear. In most situations, the link utilization is
over 90%, and the TCP throughput is improve by 100%. We
have also noticed that some factors and events likely impose
the negative impact on performance of our improved scheme
under particular environment and configuration, such as the
large propagation delay between NodeB and our proxy, and
the loss of feedback signaling caused by unreliable UDP. In
future work, we will try to design the robust control algorithm
Fig. 7. Evolutions of various key parameters.
against these disturbances. In addition, it is challenging and
attractive for us how to build an analytical model to evaluate
TABLE I
S TATISTIC RESULTS
the throughput of TCP and our TCP proxy over the wireless
link with bandwidth oscillation.
Throughput Avg. Throughput
T upper bound (pkts/s) Link utilization (%) ACKNOWLEDGMENT
(pkts/s) TCP Proxy TCP Proxy
2s 119.2 69.5 118.8 58.26 99.70 This work is supported in part by the National Natural Sci-
200ms 124.0 60.0 115.1 48.32 92.82 ence Foundation of China (NSFC) under Grant No.60573122
20ms 123.82 18.6 115.5 15.04 93.29
and 60773138; the National Grand Fundamental Research
Program of China (973) under Grant No.2003CB314804.
 150
average throughput should be 0 bw(t)dt/150 = 124.0 R EFERENCES
packets/s, and the actual throughput is 115.1 packets/s in our
[1] H. Balakrishnan, V. Padmanabhan, S. Seshan, M. Stemm, and R. H. Katz,
TCP Proxy, however, it is 60.0 packets/s in the standard TCP. “A comparison of mechanisms for improving TCP performance over
Thus the link utilization is 92.82% and 48.32% respectively. wireless links,” IEEE/ACM Trans. Networking, vol. 5, pp. 756-769, 1997,
The former is about twice as the latter. To demonstrate the [2] A. Chockalinggam and M. Zorzi, “Wireless TCP performance with link
layer FEC/ARQ,” in Proc. IEEE ICC’99, pp. 1212-1216, June 1999.
impact of the bandwidth oscillation frequency on TCP perfor- [3] M. Yavuz and F. Khafizov, “TCP over wireless links with variable
mance, we change the period of random sequences modulating bandwidth,” in Proc. IEEE Vehicular Technology Conference (VTC’02
the bandwidth oscillation. Setting T =2s and 20ms, we repeat Fall), Sept. 2002.
[4] M. C. Chan and R. Ramjee, “TCP/IP performance over 3G wireless links
the above simulations, and record data and calculate the link with rate and delay variation,” in Proc. ACM Mobicom, pp. 71-82, 2002.
utilization. The results are presented in Table 1. In the standard [5] M. C. Chan and R. Ramjee, “Improving TCP/IP performance over third
TCP protocol, the speed of bandwidth variation severely generation wireless networks,” in Proc. IEEE INFOCOM 2004.
[6] F. Khafizov and M. Yavuz, “Running TCP over IS-2000,” in Proc. IEEE
affects the TCP performance. As T =20ms, the link utilization ICC 2002.
is only 15.04%. For our TCP Proxy, the link utilization also [7] A. Gurtov and S. Floyd,“Modeling wireless links for transport protocols,”
decreases as the bandwidth oscillation frequency increases; ACM Computer Commun. Rev., vol. 34, no. 2, pp. 85-96, Apr. 2004.
[8] A. Baker and B. R. Badrinath, “I-TCP: indirect TCP for mobile net-
however, it is always over 90% and acceptable. Generally, works,” in Proc. 15th International Conference on Distribution Comput-
our TCP proxy can improve TCP performance by close to ing Systems 1995.
100% (115-60/60=0.918, T =200ms), and even 500% (115.5- [9] M. Kohlwes, J. Riihijarvi, and P. Mahnen, “Measurement of TCP per-
formance over UMTS networks in near-ideal conditions,” in Proc. IEEE
18.6/18.6=5.20, T =20ms) in the extreme case. Vehicular Technology Conference (VTC’05), vol. 4 pp. 2235-2238, May
2005.
VI. C ONCLUSION [10] E. Altman, F. Baccelli, and J. C. Bolot, “Discrete-time analysis of
adaptive rate control mechanisms,” High Speed Networks and Their
The various link adaptation techniques employed by HS- Performance, Perros and Viniotis, eds., North Holland, 1994, pp. 121-
DPA provide higher peak data rates, but they also aggra- 140.
[11] W. Wei, C. Zhang, H. Zang, J. Kurose, and D. Towsley, “Inference
vate bandwidth variations, which will severely deteriorate and evaluation of split-connection approaches in cellular data networks,”
TCP throughput. Comparing with other network technology, in Proc. Passive and Active Measurement Conference 2006, Adelaide,
HSDPA network has a unique advantage, namely NodeB Australia, Mar. 2006.
[12] C. Chatfield, The Analysis of Time Series: An Introduction, 6th edition.
determinately knows the value of bandwidth assigned to each CRC Press LLC, 2004.
user, which is very helpful for optimizing TCP performance. [13] G. Strang, Linear Algebra and its Applications, 2nd edition. Academic
Motivated by these understanding, we propose a performance Press, 1980.