Vous êtes sur la page 1sur 5

Simulation-Based Performance Analysis of HSDPA for UMTS Networks

Aun Haider, Richard Harris and Harsha Sirisena

Abstract— This paper investigates the performance of TCP switching/routing of calls and the User Equipment (UE).
over High Speed Downlink Packet Access (HSDPA) link in UTRAN consists of a Radio Network Controller (RNC)
the third generation Universal Mobile Telecommunication Sys-
tem (UMTS). The throughput performance of TCP has been
considered with both round robin and maximum carrier-to-
interference scheduling schemes. A simulation based approach
is adopted to determine the total cell throughput, both under
ideal and Rayleigh fading conditions. Simulation results show
that, HSDPA can effectively provide high throughputs, but with
fairness depending upon the scheduling mechanism. RNC
Node-B
SGSN GGSN IP Networks
I. INTRODUCTION
The revolution of wireless networking continues to un- Gn
RNC Core Network
fold with the gradual emergence of WCDMA-based third
UE Node-B
generation (3G) cellular technologies, such as Universal Iub Gi

Mobile Telecommunications System (UMTS) [1]. The range Utran


of offered services in such systems has already extended Uu
IuPS
from basic speech telephony to multimedia and interactive
data transfers. In such an environment one can expect that a Fig. 1. An overview of UMTS Network Topology
major portion of the overall traffic will be carried by TCP/IP.
Thus, particular attention must be paid to the performance and Node-B which acts as a base station. Node-B contains
of TCP/IP over 3G wireless networks. functionalities for fast link adaptation, fast hybrid ARQ
TCP/IP performance over wireless networks has been (HARQ) and fast scheduling. The core network consists of
extensively studied in the last decade. It has been found that two basic nodes: Serving GPRS Support Node (SGSN) and
wireless link losses have an adverse impact on performance Gateway GPRS Support Node (GGSN), where the former
as TCP/IP cannot distinguish them from congestion losses. node is connected to the RNC via the IuPS interface and
These findings have been one of the main motivations for the latter node provides interworking with external packet
the use of extensive local retransmissions in 3G networks. switched networks over the Gi interface. Both SGSN and
In this paper we will restrict ourselves to UMTS based 3G GGSN have already been defined for the General Packet
networks only. Radio Switch (2.5G) extension to GSM networks. The UE
The rest of the paper is organised as follows. An overview is connected to UTRAN over the UMTS radio interface Uu.
of UMTS including architecture, protocol stack, channel
capacities and scheduling mechanisms has been given in B. Protocol Stack
section II. Related work is presented in section III. Details
The protocol stack for UMTS in the transmission of user
of network model and simulation results are presented in
plane data generated by TCP/IP and UDP based applications
section IV. Finally, conclusions are drawn in section V.
is shown in Fig. 2, [2]. TCP/IP protocol suite and applica-
II. A N OVERVIEW OF UMTS tions are located both at the UE and the end host. UDP and
A. Network Architecture IP are used as a means to transport traffic and signalling
information among GGSN, SGSN and RNC. The network
A reference architecture for UMTS in packet switched layer, radio resource and radio link control are divided into
mode is given in Fig. 1, [2]. It builds on the well known Control (C-) and User (U-) planes. Information sent and
architecture of second generation networks. The major func- received by the UE, such as voice calls or internet packets,
tions of the system can be grouped into the UMTS Radio is transported over the U-plane, whereas the C-plane is used
Access Network (UTRAN) which handles all radio related for signalling.
functionalities, the core network which is responsible for Packet Data Convergence Protocol (PDCP) exists only in
Aun Haider and Richard Harris are with Institute of Information Sci- the U-plane and its main function is header compression
ences and Technology, Massey University, Palmerston North, New Zealand which improves the spectral efficiency for transmitting IP
a.haider@massey.ac.nz, r.harris@massey.ac.nz packets. The Radio Link Control (RLC) layer can operate
Harsha Sirisena is with Deptartment of Electrical and Computer
Engineering, University of Canterbury, Christchurch, New Zealand in three different modes, viz: Transparent Mode (TM), Un-
sirisehr@elec.canterbury.ac.nz acknowledged Mode (UM) and Acknowledged Mode (AM).
In AM, Automatic Repeat Request (ARQ) is used for error Automatic Repeat Request (HARQ), Higher peak rates (up to
correction until data is received correctly. In both TM and 14.4 Mbps in downlink) and moving the scheduling functions
UM, no retransmission protocol is used and data delievery from Radio Network Controller to Node B (thus splitting
is not guaranteed. Erroneous protocol data units may be the MAC layer). HSDPA defines a new WCDMA channel,
discarded or marked. The RLC protocol runs in both RNC called high-speed downlink shared channel (HS-DSCH), for
and UE, where it implements regular data link functionality downlink communication to mobile cellular devices.
over the WCDMA interface and provides segmentation and
D. Packet Scheduling
retransmission services, [2].
The Medium Access Protocol (MAC) is active at both In the HSDPA context, the process of scheduling refers to
the UE and RNC entities and it maps the logical channels the process of allocation of transmitter time and power (i.e. at
to transport channels. The Physical Layer (PHY) maps the base station, Node-B) to the randomly time-varying mobile
transport channels into physical channels and also provides data connections (i.e. cellular mobile users, UE). In such a
functions for modulation, demodulation, spreading, synchro- case, the simplest type of scheduling (channel non-adaptive),
nization, interleaving, forward-error correction and soft han- is to allocate the access time fairly to all users, and this is also
dover. known as Round Robin (RR) Scheduler. Although RR is fair
in time allocation, its performance is not very satisfactory in
fading environments, [3].
APPL
Another scheduling approach can be based on transmit-
TCP UDP ted power by the mobile users, as it is representative of
IPv4/IPv6 IPv4/IPv6 channel conditions. In this case Node-B can compare the
PDCP PDCP
GTP−U
UDP
GTP−U
UDP UDP
GTP−U
UDP
power transmitted by each mobile users and match it to the
RLC RLC IP IP IP IP respective packet waiting in the queue for transmission. Thus,
MAC MAC AAL5 AAL5 L2 L2
Node-B will transmit a packet to the user which has the
PHY PHY FP FP ATM ATM
best channel conditions and so on (channel adaptive). This
L1 L1
TN TN approach is known as Best Channel First (BCF) or Maximum
UE Node B RNC SGSN GGSN Carrier-to-Interference ratio (C/I). Maximum C/I is unfair as
Uu Iub IuPS Gn Gi it favours the users who are closer than users who are further
Fig. 2. UMTS Protocol Stack
away from Node-B. In order to have a compromise between
RR and Maximum C/I, Fair-Channel Dependent Scheduling
(FCDS) has been proposed in [4].
C. UMTS Channel Capacities In our simulation experiments, the performance of TCP
For the UMTS release 99 downlink, data can be trans- is determined under two extreme conditions of fairness by
ferred through the shared Forward/Random Access Channel employing the RR and Maximum C/I scheduling schemes.
(FACH/RACH) as well as on the Dedicated Channel (DCH). It can be noted that FCDS can be made to behave as RR or
The DCH can have a high throughput but it also requires Maximum C/I, by an appropriate choice of a parameter α,
a setup time of the order of 250 ms, which may be too [4].
long for short data transfers. On the other hand, the shared The thoroughput at UE depends upon the scheduling
channel has low setup time but also low throughput. Table scheme employed at Node-B. In [5] it has been stated that
I presents the capacities of the different channels in UMTS. the total cell throughput of a HSDPA system having a total
Release 5 of the UMTS-standard incorporates High Speed of Nu users with a mean bit rate of Ri is given by
Nu
!
TABLE I X
T =E Ri , (1)
UMTS C HANNEL C APACITIES
i=1
where for RR, Ri is given by
Transport Channel Types
Common Channel Dedicated Channel 1 X W (N log 2(M )τ )m,i
RACH FACH DCH
Ri = km . (2)
Nu m SF Ns,i
Bit rate TTI Bit rate TTI Uplink Downlink
(Kb/s) (ms) (Kb/s) (ms) Bit rate TTI Bit rate TTI In (2) W is a chip rate, SF is spreading factor, Ns,i is the
(Kb/s) (ms) (Kb/s) (ms)
32 10 32 10 64 20 64 20 number of transmissions for user i due to HARQ, M is
64 20 128 20 the modulation order and τ is code rate, [6] and [5]. It has
64 20 384 10 been mentioned that km varies with the position of UE. For
384 10 2000 10
maximum C/I scheduling, the expression for Ri is the same
as in (2) but multiplied to the probability that a Transmission
Downlink Packet Access (HSDPA), which is a significant Time Interval (TTI) is allocated to user i, for which no simple
advance that can offer Mobile Broadband services and is also or closed form formula has been derived.
refered to as 3.5G. HSDPA has been based on four funda- In [7] anlytical expressions for the RLC part of total
mental principles: Adaptive Modulation and Coding, Hybrid round trip time and throughput have also been derived. The
expression for throughput involves the round-trip time for As background traffic, there are 5 UDP connections, each
the wired part, for which no expression is given. Thus, running CBR traffic. Each of the 20 UEs are connected to
to determine the throughput at UEs, a simulation based Node-B through an acknowledged mode HSDPA channel as
approach IS adopted in this paper. in [2]. The Node-B is connected to Wired Node 1 through
RNC, SGSN, GGSN and wired node 2. Thus each UE has
III. R ELATED W ORK an end-to-end TCP connection with wired node 1 as shown
A comparison between WCDMA Release 99 and HSDPA in Fig. 3.
has been provided in [8]. The performance of WiBro, a new The simulations have been run both under ideal and
Korean standard, and HSDPA has been provided in [9]. It has Rayleigh fading environments. To simulate the ideal con-
been found that WiBro performs better in multipath fading ditions we use a trace, in which there are no Channel errors
due to the use of Orthogonal Frequency Division Mutiple and have a fixed Channel Quality Indicator (CQI) of 14, [2].
Access and a Cyclic prefix, whereas HSDPA was more For Rayleigh fading we use UEs that are 700 m away from
robust in Doppler shift fading due to its shorter Transmission Node-B and moving with velocity 3Km/h. The trace files
Time Interval. Also it has been anticipated that both WiBro under different environments and fading conditions can be
and HSDPA will compete fiercely with each other in the generated using Matlab code available at [2].
broadband mobile market. The simulated network with link capacities, is shown in
In [10] it has been reported that TCP throughput is close Fig. 3. The link between Base station (Node-B) and Radio
to theoretical values and round trip time is stable. Also it Network Controller (RNC) is using a Droptail queue with
has been mentioned that no performance benefit has been arbitrarily large size of queue buffer, so that there is no packet
obtained with modifications in the TCP retransmission timer, loss due to overflow. The parameters of the exponential
such as Eifel. The results presented need to studied further
both in real environments and simulations. HSDPA
622 Mb/s, 10 ms, 10 Mb/s, 35 ms,
CHANNEL
Using a quasi-static simulator the overall throughput and Droptail Droptail
622 Mb/s, 0.4 ms, 10 Mb/s, 15 ms,
delay performance of 1XTREME and HSDPA has been Droptail Droptail
provided in [11]. The performance has been bounded by
Maximum C/I and Round Robin schedulers. The link and
network layer performance aspects of WCDMA with HSDPA Mobile Units
NODE-B RNC SGSN GGSN Wired Node 2 Wired Node 1
UE
have also been studied in [12]. Six prototype packet sched-
ulers have been considered and inherent tradeoff among cell Fig. 3. Simulated UMTS Network
capacity and user fairness has been illustrated. Under differ-
ent environments and traffic aspects, flow level performance
ON/OFF and CBR traffic are the same as the default values
of HSDPA, has also been investigated in [13].
in the ns-2 simulator (i.e. burst and idle times of 0.5 s each
In [6] and [5] the authors have developed an analyt-
with a rate of 64 Kb/s for exponential traffic; data rate of
ical model to evaluate the impact of TCP on UMTS-
448 Kb/s for CBR traffic). Data transfer on the Internet has
HSDPA capacity and have presented a method (maximize
been modelled as the heavy tailed Pareto distribution with
Nc log 2(M )τ , where Nc is number of codes available, M
CDF given by:
is the modulation order and τ is code rate) to maximize the
 α
number of users in a cell i.e. cell capacity. The analytical k
model has been validated by using both Monte Carlo and F (x) = P [X ≤ x] = 1 − , (3)
x
NS based simulations.
where k is the minimum value X and α is a shape factor.
IV. N ETWORK M ODEL In [14], it has been found that the shape factor, α, ranges
We consider a single cell in which one base station, i.e. from 1.04-1.14. In our simulations we have used Pareto
Node-B, is serving n mobile users on a HSDPA channel distribution based ON/OFF traffic carried on TCP with burst
using Hybrid CDMA/TDMA transmission scheme. Further, time and idle time each having a mean of 500ms, data rate
assume that at a given timeslot t, there are m, m ≤ n of 64 Kb/s and shape factor of 1.09.
active mobile users who are competing with each other for The throughput obtained at each of the 15 UEs, under ideal
resources at Node-B. The packets to be transmitted have been conditions, has been plotted in Fig. 4. Also, for both RR and
enqueued at Node-B. For simulation purpose we employ Max C/I scheduling, throughputs has been plotted in Figs.
Enhanced UMTS Radio Access Network Extensions (EU- 5 and 6, respectively. For ideal conditions, the throughput
RANE) to ns-2 simulator, in which handover functionality of each UE achieves a steady value between 60 and 65
has not been implemented, [2]. Kb/s after 10 s of simulations. In the case of Rayleigh
We simulate a scenario in which 20 mobile users are fading with RR scheduling, there is an intial transient in
downloading data on a HSDPA link from Node-B using throughput during the period between 0 to 40 s, which after
TCP with selective acknowledgements and UDP. In this 80 s decreases smoothly as UEs move away from Node-B.
scenario, the number of TCP connections running FTP, For maximum C/I scheduling the transient state in throughput
exponential and Pareto (ON/OFF) distributed traffic is 5 each. lasts for a period of 80s. In Fig 6 it can be seen that UE 15
gets much less throughput, thereby depicting the unfairness
700
caused by the maximum C/I scheduling policy. UE 1
UE 2
The total cell throughput obtained by all UEs, both for UE 3
UE 4
600 UE 5
RR and maximum C/I, has been plotted in Fig. 7. It shows UE 6
UE 7
that maximum C/I will generate a higher value of total cell 500
UE 8
UE 9
UE 10
throughput as compared to RR. Queue size variations in the

Throughput (Kb/s) at UE
UE 11
UE 12
link between Node-B and RNC have also been plotted in 400
UE 13
UE 14
UE 15
Figs. 8 and 9, respectively. These show that RR has slightly
higher queue size than maximum C/I. The variations in end- 300

to-end delay for TCP packets have been plotted in Figs. 10


200
and 11.
70 100

60 0
0 20 40 60 80 100 120 140 160
Time (s)

50
Fig. 6. Throughput for 15 UE’s under Rayleigh fading with max C/I
Throughput (Kb/s) at UE

scheduling.
40

UE 1
30 UE 2
UE 3
UE 4
UE 5 2500
UE 6 RR Sceduling
20 UE 7 Max C/I Scheduling
UE 8
UE 9
UE 10
10 UE 11
UE 12 2000
UE 13
UE 14
UE 15
Total cell throughput (Kb/s)

0
0 10 20 30 40 50 60 70 80 90 100
Time (s) 1500

Fig. 4. Throughput for 15 UE’s under ideal conditions.


1000

140
UE 1 500
UE 2
UE 3
UE 4
120 UE 5
UE 6
UE 7
UE 8 0
100 UE 9 0 20 40 60 80 100 120 140 160 180
UE 10 Time (s)
Throughput (Kb/s) at UE

UE 11
UE 12
UE 13
80 UE 14
UE 15
Fig. 7. Total cell throughput with RR and Max C/I Scheduling

60

40 25
Inst queue with RR

20

20
Instantaneous queue size (Packets)

0
0 20 40 60 80 100 120 140 160
Time (s)
15

Fig. 5. Throughput for 15 UE’s under Rayleigh fading with RR scheduling.

10

V. C ONCLUSIONS
In this paper, a simulation based approach has been 5

adopted to evaluate the performance of HSDPA in 3G UMTS


networks. It has been found that, under ideal conditions,
0
throughputs of all 15 UEs converge quickly (within around 0 20 40 60 80 100 120 140 160
Time (s)
20 s) to a steady-state value of 60 Kb/s, whereas in the case
of Rayleigh fading conditions the throughput convergence Fig. 8. Instantaneous queue variations (between Node-B and RNC) with
depends upon the scheduling policy. Round robin scheduling RR scheduling.
is shown to distribute channel resources fairly, but also to
yield a larger queue size, i.e. buffer occupancy, in the link
16
Inst queue with Max C/I between Node-B and RNC. For higher total cell thorugh-
14
put, maximum C/I scheduling is preferred, though it will
cause severe unfairness by providing only a small channel
12 bandwidth to some of the UEs. The variation in end-to-
Instantaneous queue size (Packets)

end packet delay is shown to be lower for maximum C/I


10
then RR, as some UE throughputs are very low, due to
8 unfair distribution of channel bandwidth, in the former case.
Performance evlaulation of HSDPA with other scheduling
6
mechanisms, such as proportional fair throughput, is left for
4
a future work. Another direction of future work may be to
employ AQM controllers (RED, PI and PID) between Node-
2 B and RNC.
0
0 20 40 60 80 100 120 140 160
R EFERENCES
Time (s)
[1] 3GPP, “http://www.3gpp.org/,” .
[2] Seacorn, “Deliverable d3.2v2, end-to-end network
Fig. 9. Instantaneous queue variations (between Node-B and RNC) with model for enhanced umts,” Available at http://www.ti-
Max C/I Scheduling. wmc.nl/eurane/D32v2Seacorn.pdf.gz, http://www.ti-
wmc.nl/eurane/eurane user guide 1 6.pdf and EURANE code
available at http://www.ti-wmc.nl/eurane.
[3] Lei Zan, Geeret Heijenk, and Magda El Zarki, “Fair & Power-Efficient
Channel-Dependent Scheduling for CDMA Packet Networks,” Pro-
62.6
ceedings of International Conference on Wireless Networks (ICWN),
62.4
2003, June 2003, Available at http://wwwhome.cs.utwente.nl/ heijenk/.
[4] Lei Zan, Geeret Heijenk, and Magda El Zarki, “A Real-Time Traffic
62.2 Scheduling Algorithm in CDMA Packet Networks,” Proceedings of
End-to-end delay in TCP packets (ms)

14th IEEE International Symposium on Personal, Indoor and Mobile


62
Radio Communications PIMRC 2003,, September 2003, Beijing,
61.8
China.
[5] Mohamad Assaad and Djamal Zeghlache, “Cross-Layer Design in
61.6 HSDPA System to Reduce the TCP Effect,” IEEE Journal on Selected
Areas in Communications, vol. 24, no. 3, pp. 614–625, March 2006.
61.4
[6] Mohamad Assaad and Djamal Zeghlache, “How to minimize the TCP
61.2
effect in a UMTS-HSDPA system ?,” Wireless Communications and
Mobile Computing, vol. 5, pp. 473–485, 2005.
61 [7] Olivier De Mey, Laurent Schumacher, and Xavier Dubois, “Opti-
mum Number of RLC Retransmissions for Best TCP Performance in
60.8 UTRAN,” Proceedings of IEEE International Symposium on Personal
Indoor and Mobile Radio Communications PIMRC, 2005.
60.6
0 20 40 60 80 100 120 140 160 [8] Che-Sheng Chiu and Chen-Chiu Lin, “Comparative Downlink Shared
Time (s) Channel Performance Evaluation of WCDMA Release 99 and HS-
DPA,” Proceedings of the 2004 IEEE International Conference on
Fig. 10. End-to-end delay in TCP packtes, with RR scheduling. Networking, Sensing and Control, pp. 1165–1170, March 2004.
[9] Simon Shin, Chan koo Kang, Joung-Cheol Kim, and Se-Hyun Oh,
“The Performance Comparison between WiBro and HSDPA,” Second
International Symposium on Wireless Communication Systems, pp.
346–350, September 2005.
62.4 [10] Martin Kohlwes, Janne Riihijarvi, and Petri Mahonen, “Measurement
of TCP Performance over UMTS Networks in Near-Ideal Conditions,”
62.2 Proceedings of IEEE Vehicular Technology Conference, VTC 2005,
vol. 4, pp. 2235–2239, May-June 2005.
62
[11] Robert Love, Amitava Ghosh, Robert Nikides, Louay Jalloul, Mark
End-to-end delay in TCP packets (ms)

Cudak, and Brian Classon, “High Speed Downlink Packet Access


61.8
Performance,” Proceedings of IEEE Vehicular Technology Conference,
61.6
vol. 3, pp. 2234–2238, May 2001.
[12] Troels E. Kolding, Frank Frederiksen, and Preben E. Mogensen, “Per-
61.4
formance Aspects of WCDMA Systems with High Speed Downlink
Packet Access (HSDPA),” Proceedings of IEEE Vehicular Technology
61.2 Conference, vol. 1, pp. 477–481, September 2002.
[13] Hans van den Berg, Remco Litjens, and Joost Laverman, “HSDPA
61 Flow Level Performance: The Impact of Key System and Traffic
Aspects,” Proceedings of ACM MSWiM’04, pp. 283–292, October
60.8 2004, Venezia, Italy.
[14] Bruce A. Mah, “An Empirical Model of HTTP Network Traffic,”
60.6
0 20 40 60 80 100 120 140 160
Proceedings of IEEE INFOCOM’97, vol. 2, pp. 592–600, April 1997.
Time (s)

Fig. 11. End-to-end delay in TCP packtes, with max C/I scheduling.

Vous aimerez peut-être aussi