Preventive and Reactive based TNL Congestion

Control Impact on the HSDPA Performance

Thushara Weerawardane1, Andreas Timm-Giel1, Gennaro C. Malafronte3, Durastante Gianluca3, Stephan Hauth2,
Carmelita Grg1
1T

TZI ComNets, University of Bremen, Otto-Hahn-Allee NW1, 28359 Bremen, Germany, {tlw/atg/cg}@comnets.uni-bremen.de
2
Nokia Siemens Networks GmbH & Co. KG, Lise-Meitner-Str. 7/2, 89025 Ulm, Germany, {stephan.hauth@nsn.com}
Nokia Siemens Networks S.P.A., Via Monfalcone 1, Cinisello Balsamo, Italy, {gennaro.malafronte/gianluca.durastante}@nsn.com

AbstractHigh Speed Downlink Packet Access (HSDPA)

is an extension of the Universal Mobile Telecommunications
System (UMTS) technology of 3GPP Rel-99, with the
objective to increase the data rate and reduce the latency in
the downlink. The main focus of the work presented is to
analyse the effect of congestion at the Iub interface on the
HSDPA performance. The data flows should be adequately
controlled in order to avoid congestion in the transport
network. The 3GPP (3rd Generation Partnership Project)
Rel. 5 specifications highlight two congestion detection
mechanisms which are based on the frame sequence
number (FSN) and the Delay Reference Time (DRT) fields
of HSDPA data frame. In addition to these, a third
congestion detection mechanism based of Checksum of
HSDPA data frame is considered. This paper discusses a
congestion control scheme deploying all three congestion
detection methods. It is shown, that a congestion control
algorithm can effectively work using these congestion
detection triggers and can control the offered load to the
transport network. The simulation results presented in this
paper confirm that the congestion in the transport network
can be avoided, and hence the performance of HSDPA
network can be significantly improved in all aspects.
KeywordsHSDPA, Congestion Detection, Congestion
Control

I. INTRODUCTION
HSDPA is an upgrade to the 3GPP Rel-99 UMTS
technology meeting the ever increasing user requirements.
It is an asymmetric solution that enables higher downlink
throughput to the end user. Several enhancements to
WCDMA R99 were considered in 3GPP release 5 [1].
Moving the scheduler from the RNC (Radio Network
Controller) to the NodeB and reducing the Transmission
Time Interval (TTI) from 10ms to 2ms allows adapting
better to the varying radio channel. The Adaptive
Modulation and Coding (AMC) can adapt the modulation
rank and coding rate to the current radio channel
condition of each user. With a constant transmit power,
the effective data rate can so be significantly enhanced
for users with good radio channel conditions. The other
prominent improvement of the physical layer is the use
of Hybrid ARQ (HARQ), where the soft bits from
retransmission requested by the UE are combined with
soft information from the original transmission prior to
the decoding [2] leading to an improved BER
performance.
In addition to the physical layer improvements, currently
the research is focused on enhancements and
performance measures on Iub interface [3], [4], [5] and
[6]. From the operator point of view, the Iub is usually a
scarce resource. The HSDPA transport channel, HSDSCH [1], [2] is controlled by the MAC-hs scheduler
978-1-4244-1645-5/08/$25.00 2008 IEEE

which is a new entity located in the Node B. Packet switched

traffic via HS-DSCH is scheduled in the Node B and then
transmitted over the air interface. Moving the scheduler to the
Node B enables more efficient implementation of the
scheduling by allowing the scheduler to work with the most
recent channel information. The required data for radio
channels should arrive with low delay at the Node B side in
order to minimize the data buffering in the MAC-hs. The
Transport Network Layer (TNL) is optimized by deploying
suitable flow control and congestion control techniques
minimising the Iub link bandwidth requirements, providing
better QoS to the end user and also reducing the operator costs
[4], [5], [6] and [8]. Compared to Rel-99, HSDPA uses two
buffering points at the Node B and at the RNC. Since the
transport capacity is limited and the demand of the air interface
varies rapidly according to the channel condition, the data flow
over the Iub interface should be adapted accordingly. The
credit based flow control scheme quantifies the required air
interface capacity and informs the RNC as a resource
allocation for each user flow separately. However, when the
demanded user traffic over the air interface is greater than the
available Iub capacity, the congestion occurs at the UTRAN
transport. Due to Iub congestion, the packet loss probability
increases significantly, causing more retransmissions at higher
layers such as RLC and TCP. Hence the offered load is further
staggered leading the system into congestion collapse and
wasting network resources. To avoid such circumstances, a
proper congestion control scheme is required in addition to the
credit based flow control mechanism.
A credit based flow control mechanism has been developed in
[4] to optimise the Iub utilisation while providing required
QoS to the end user in the HSDPA network. It smoothes down
the HSDPA traffic and reduces the burstiness over the Iub
interface. The implemented credit allocation algorithm is based
on the periodically updated provided bit rate (PBR) of the
user-specific priority queues in the Node B.
As mentioned above when the Iub link is overloaded,
congestion occurs and cannot be protected by the flow control
mechanism and a proper TNL based congestion control
mechanism is required. 3GPP has introduced two basic
mechanisms working at the FP layer (i.e. transport media
independent) to perform Iub congestion detection (see TS
25.435): frame loss detection by means of FSN (Frame
Sequence Number) supervision and delay build up detection
by means of DRT (Delay Relative Time) supervision. In
addition to these, a third congestion detection mechanism
based of Checksum of HSDPA data frame is considered.
The remainder of this paper is structured as follows. The
preventive and reactive congestion detection mechanisms are
discussed in details in the second and Third chapters. Next, the
congestion control procedures which are activated upon the
congestion detection triggers are described in details. Then the
2296

simulation model and simulation scenarios are presented

in following chapter. Finally, the simulation results and
analysis are presented along with the conclusion at the
end of the paper.

i=2

i =1

R(t i ) = (t i ) = TNLDELAY (i, i 1), i 1 .

II. REACTIVE CONGESTION DETECTION

The reactive Iub congestion detection mechanism
detects the congestion upon detection of packet losses at
the transport network. Since the TNL uses the ATM as
the transport technology between Node B and RNC,
small cell losses can occur when the system is
overloaded. This either leads to corrupted frames or loss
of complete frames, the later due to the bursty losses.
When the header CRC check fails, the packet cannot be
identified and is discarded completely. In many situations,
the FP layer receives frames with a valid header but
invalid payload. Such errors are identified as the payload
CRC errors. The most common frame errors can be
categorised and described as follows.
1. Missing a last segment or tail of the FP frame
When the last segment or the tail of the FP PDU is lost,
the receiver waits until it receives the last segment of the
next frame to be reassembled. This results in creating a
large frame with an invalid CRC during the reassembly
process. Both original frames are lost.
2. Missing any segment except the tail of a frame
The receiver can reassemble a FP frame having an
invalid CRC due to a missing segment or due to an
insertion of a foreign segment. Since the tail of the frame
is preserved, the next frame can be detected again.
3. Loss of complete frames
When a burst of cell losses occurs at the transport
network, one FP frame or several FP frames can be lost.
These kind of bursty losses are very common for HSDPA
PS networks when congestion persists and are detected
monitoring the FSN.
Another important reason besides Iub congestion for
missing or invalid frames is given physical layer errors;
Such physical layer errors are filtered out by the Iub CD
algorithm by means of a probabilistic discriminator.
The error type 1 (missing a last segment or tail of the FP
frame) and 2 (missing any segment except the tail of a
frame) can be easily identified by the FP payload CRC
check. The type 3 (Loss of complete frames) can be
identified by using a frame sequence number (FSN).
III. PREVENTIVE CONGESTION DETECTION
The preventive congestion detection is in charge of
detecting Iub packet losses before they occur. The
preventive DRT based delay build-up algorithm monitors
the FP PDU delay variation for the correctly received FP
frames through the transport network for each MAC-d
flow. The figure below shows the delay variation for
several FP PDU transmissions with respect to the RNC
and Node B reference counters.
According to the Figure 1, the delay build-up algorithm
monitors the TNL delay variations for each arrival of an
FP frame at the Node B. For example, the delay variation
between frame i and i-1 is given in the following formula.

( t i ) = TNL DELAY (i, i - 1), i 1

where ( t i ) is the delay variation between i and i-1.
The accumulated delay variation is given by R(ti),

Fig. 1: FP PDU delay build-up behaviors through TNL network

Once the accumulated delay variation exceeds a certain

threshold within a certain time window, a congestion is
assumed and congestion control is initiated.
IV. CONGESTION CONTROL
The functional block diagram of the flow control, congestion
detection and congestion control are shown in Fig. 2.

Fig. 2: EFC and congestion control mechanism in HSDPA network

Both congestion detection and congestion control algorithms

are implemented in a cascade with the credits based flow
control mechanism which is used to estimate the current air
interface capacity in the Node B for HSDPA [4]. The
congestion detection is triggered upon each arrival of an FP
PDU at the Node B. The figure also shows the two inputs to
the congestion control algorithm: one is provided by the FC
module and the other is provided by the congestion detection
module. The FC and the congestion detection algorithms work
independently and provide information to the congestion
control module promptly.
The MAC-d flow can be either in congestion state or in flow
control state. Therefore, the system can be represented by a
flow state machine as shown in the following figure.

Fig. 3: HSDPA MAC-d flow state machine

When MAC-d Flow is in the flow control state Capacity

Allocation Messages (CA) are sent to RNC periodically [4] by
updating the allowed credits. Once the congestion detection
2297

events are triggered by the detection module the MAC-d

flow goes to Congestion Control State (CC) by sending
the CA message to the RNC. At the congestion control
state, the MAC-d flow is controlled by the AIMD (a,b)
algorithm [9], [10]. The credits allocation is done on top
of the flow control credits and the amount of credit
reduction is done according to the severity of the
congestion trigger which is received by the CC module.
The AIMD (a, b) rate control mechanism is used to
reduce the rate by b100% at each congestion detection
and subsequently the rate is increased by a credits at
the end of user defined time interval in the recovery
phase. The congestion control module always compares
both FC credits and CC credits, and sends the lowest
between them to the RNC in a CA message. The
condition for the MAC-d flow to exit the CC state is that
the value of credits calculated by AIMD (a,b) either
comes back to the value of credits assigned to the MAC-d
Flow when it entered the CC state or it is exceeded by the
flow control credits.

idea of these two configurations with two different scenarios is

to analyse the effect of the congestion control for two different
offered load situations.
The effect of the credit based flow mechanism was tested and
presented in [4]. The key results presented in this paper are
used to evaluate the effect of the TNL congestion control
algorithm on the performance of the HSDPA network.

V. SIMULATION MODEL AND SIMULATION

SCENARIOS

Link level as well as the end user level performance is

analysed. The ATM throughput statistics explain the utilisation
at the transport network, whereas the IP throughput statistics
show the performance at the end user entities. The delay
statistics and the loss ratio statistics show the impact of the
congestion control scheme on the TNL network.

The comprehensive HSDPA simulation model is

developed under the OPNET simulation environment.
The simulation model is specially designed for the TNL
feature analysis, but allows in addition performance
analysis of RLC, TCP and application layer on a per user
level. More details about the model design and overview
are discussed in [3], [4]. In the next sections the traffic
models and parameters used for the simulation and
analysis of the Congestion Control are presented.
The most common type of HSDPA packet traffic consists
of web and FTP traffic which uses the TCP
(Transmission Control Protocol) as a reliable end-to-end
transmission protocol. The web traffic model which is
defined by the ETSI standards [1], [7] is selected to
evaluate the performance of HSDPA traffic under
moderate network load. The FTP traffic model is used to
overload the network therefore, it provides the worst case
traffic scenarios to analyse the impact of the above
features on the end user performance as well as the
overall network performance. Under this traffic
configuration, all users in the cell download a large file
and utilize the network resources up to the maximum
available capacity. The parameters of the web traffic
model defined by ETSI and the FTP traffic model are
given in the Table I. Both ETSI and FTP traffic models
are adapted as application layer traffic models which
include the TCP protocol adaptation as well. The
windows 2000 based TCP configuration that uses the
TCP Reno version, is used for all simulations. Therefore,
TCP congestion control functionality is included for
these analyses.
Two simulation scenarios are defined according to the
types of traffic models they use: simulation scenario 1 is
based on the FTP traffic model, whereas simulation
scenario 2 is based on the ETSI traffic model. Each
simulation scenario is simulated with two different
configurations, each with 20 UEs in the cell.
Configuration 1 is defined by using flow control and
congestion
control
mechanisms
together
and
Configuration 2 only uses the FC mechanism. The basic

TABLE I: TRAFFIC MODELS FOR HSDPA

ETSI Traffic Model Parameters
Parameters
Distribution and values
Packet call interarrival Time
Geometric distribution
IAT = 5 seconds
Packet call size
Pareto distribution
Parameters:
=1.1, k=4.5 Kbyte,
m=2 Mbyte
MPS = 25 Kbyte
FTP Traffic Model Parameters
File size
Constant Distribution
MFS = 12 Mbyte
*IAT = Interarrival Time, MPS = Maximum packet size, MTU=
Maximum Transfer Unit

A. Simulation scenario 1: FTP traffic

In the first scenario the FTP traffic model is used, which is
considered as the worst case scenario. The simulation
configuration which includes a FP payload based congestion
control scheme is labelled with_CC whereas the other
configuration which does not include any congestion control
scheme is referred to as without_CC. Both configurations
use the flow control mechanism. All simulation results are
presented and analysed separately in the next section.
ATM link throughput
Fig. 4 shows the ATM link utilisation probability
distributions. The offered traffic to the system is higher
without Congestion Control.

The offered load to

TNL increases

Fig. 4: ATM Link Throughput Probability Distribution

Since both scenarios use the same traffic models, the

additionally offered load to the TNL network is assumed to be
caused by higher layer retransmissions. To validate the
2298

assumption and exclude that the CC mechanism throttles

down the traffic unnecessarily, further results are
analysed at the higher layer protocols in the following.
throughput (kbit/sec)

Average IP Throughput (kbit/sec)

AAL2 loss ratio

The AAL2 loss ratio (LR) is shown in table II. The
AAL2 buffer discards the packets having a waiting time
longer than 50 milliseconds.

1500
1000
500
0
Throughput
(kbit/sec)

Without_CC

With_CC

654

1257

sim ulation configuratios

From the figure, it can be seen that there is a clear

advantage of using a congestion control scheme which
significantly reduces the loss ratio. The loss ratio for the
scenarios without CC is about 29% and it is reduced to
0.01% with congestion control.
FP end-to-end delay
The FP PDU end-to-end delay probability distribution
for the downlink is shown Fig. 5. The FP end-to-end
delay is measured from the point of time when an FP
PDU is sent from the FP layer in the RNC to the point of
time when it is received by the FP layer in the Node B.

Fig. 6: IP throughput

Since the achieved end user performance is clearly improved,

it can be concluded that the Iub CC algorithm provides actual
benefits in terms of end user application throughput as well as
better UTRAN transport resource utilization.
B. Simulation scenario 2 ETSI traffic
Simulation scenario 2 is tested using the moderate ETSI
based traffic model. The ATM Iub bandwidth is configured to
1 Mbit/s and the simulation duration is 2000 seconds. The
following figures show the results to evaluate the performance
of the congestion control scheme for the bursty traffic model.
ATM Link throughput
As in the previous simulation, these two simulation
scenarios show the ATM utilisation with and without the CC
algorithm.

Improved delay
performance

The offered load increases

mainly due higher layer
retransmissions
Fig. 5: FP end-to-end Delay probability distribution

Fig. 5 shows the average FP end-to-end delay which is

about 45 ms for the scenario without CC and about 10 ms
for the scenario with CC. This indicates about 77%
reduction of the FP average delay at the transport
network layer.
IP throughput
The overall average IP throughput for all users is
shown in Fig. 6. The IP throughput is measured between
end user entities; it shows the overall performance of the
UTRAN as well. The average IP throughput is about
1257 kbit/sec for CC based configuration, and about 654
kbit/sec without CC. These results confirm a significant
gain of average IP throughput for the scenario which uses
the congestion control algorithm compared to the
scenario which does not use the CC algorithm.

Fig. 7: Average ATM Link Throughput

Fig.7 shows the ATM link throughput probability

distribution for both simulation scenarios. The scenario with
the congestion control scheme shows less offered traffic to the
TNL network compared to the scenario without the congestion
control scheme. However the analysis of the AAL2 frame loss
and IP throughput shows that the application end user is
improved, as it is reported below.

2299

AAL2 Loss Ratio

Average IP Throughput (kbit/sec)

throughput (kbit/sec)

The AAL2 loss ratio is shown in table III for the ETSI
based
simulations.

As in the FTP based simulations, the figure shows a clear

reduction of losses at the transport network for the
scenario using the CC mechanism.

600
500
400
300
200
100
0
Throughput
(kbit/sec)

FP end-to-end delay

Without_CC

With_CC

380

520

sim ulation configuratios

Fig. 8 shows the probability distribution of the end-toend FP delay for both scenarios.

Fig. 9: IP throughput

VI. CONCLUSION

Improved delay
performance

This paper shows the theoretical and modelling aspects of a

preventive and reactive congestion control algorithm. The
performance of a combined congestion control along with a
flow control mechanism is presented, simulated and validated
using different HSDPA simulation environments [4].
In this simulation analysis, the simulation scenarios are
configured with the same set of network and air interface
parameters in order to validate the effect of the congestion
control feature clearly. The simulation results show that for
both traffic models (ETSI and FTP traffic) the usage of
preventive and reactive congestion control algorithms allows
to effectively control TNL congestion and minimize the
network resource waste due to higher layers retransmissions.

Fig. 8: FP End-to-End Delay Probability Distribution

The average FP delay which is about 35 ms for the

scenario without CC and about 8 ms for the scenario with
CC. This indicates about 77% reduction of the FP
average delay at the transport network layer. As in the
FTP based simulations, this indicate that the CC control
based scenario provides much better TNL delay
performance compared to the scenario without CC
control.
IP throughput
The overall average IP throughput for all users is
shown in Fig. 9. The IP throughput is measured between
end user entities. As in the FTP based simulations, the
achieved throughput for the CC based simulation is
significantly higher compared to the scenario without CC
control. This shows that end user throughput can be
enhanced by avoiding congestion at the transport network.

REFERENCES
[1]

3GPP TS 25.855 High Speed Downlink Packet Access (HSDPA):

Overall UTRAN description, 3GPP TS 25.856 HSDPA: Layer 2 and 3
aspects, 3GPP TS 25.876 Multiple-Input Multiple-Output Antenna
Processing for HSDPA, 3GPP TS 25.877 HSDPA Iub/Iur Protocol
Aspects, 3GPP TS 25.890 HSDPA: User Equipment (UE) radio.
[2] L. Zhao, T.L. Weerawardane, A. Timm-Giel, C. Grg, U. Trke, M.
Koonert: Overview on UMTS HSDPA and Enhanced Uplink (HSUPA),
Mobilfunk Technologien und Anwendungen, 11. ITG-Fachtagung, Mai
2006, Osnabrck.
[3] T.L. Weerawardane, A. Timm-Giel, C. Grg, T. Reim: Performance
Analysis of the Iub Interface in UMTS Networks for HSDPA, Mobilfunk
Technologien und Anwendungen, 10. ITG-Fachtagung, Juni 2005,
Osnabrck.
[4] T.L. Weerawardane, A. Timm-Giel, C. Grg, Thomas Reim, "Impact of
the Transport Network Layer Flow Control for HSDPA Performance",
IEE conference, Sep. 2006, Colombo, Sri Lanka.
[5] M.C. Necker and A. Weber, Impact of Iub Flow Control on HSDPA
System Performance, Proceedings of the 16th Annual IEEE International
Symposium on Personal Indoor and Mobile Radio Communications,
Berlin, Germany 2005.
[6] M.C. Necker and A. Weber, Parameter selection for HSDPA Iub flow
control, in Proc. 2nd International Symposium on Wireless
Communication Systems (ISWCS), Siena, Italy, September 2005.
[7] ETSI, Universal Mobile Telecommunications System (UMTS): Physical
layer aspects of UTRA High Speed Downlink Packet. 3GPP, TR 25.848
v3.2.0, 2001-03.
[8] P.J. Legg, Optimised Iub flow control for UMTS HSDPA, In Proc. IEEE
Vehicular Technology Conference (VTC Spring), Stockholm, Sweden,
June 2005.
[9] S. Floyd, M. Handley, J. Padhye, A Comparison of Equation-Based and
AIMD Congestion Control, ACIRI, May 2000.
[10] Van Jacobson, Lawrence Berkeley Laboratory & Michael J. Karels,
University of California at Berkeley, Congestion Avoidance and Control,
November, 1988

2300