Measurement based Analysis and Modelling of UMTS DCH Error Characteristics for Static Scenarios Wolfgang Karner, Markus Rupp Institute of Communications and Radio-Frequency Engineering Vienna University of Technology, Austria Gusshausstrasse 25/389, A-1040 Vienna, Austria Email: {wkarner, mrupp}@nt.tuwien.ac.at Abstract— For the evaluation of higher layer proto- cols, the simulation of cross-layer optimisation methods or the development of new algorithms for multimedia signal processing it is essential to know the exact error characteristics of the underlying channels. In this paper we show the analysis of the error characteristics of the UMTS DL (Down Link) DCH (Dedicated Channel) with 384kbit/s especially for static scenarios. The analysis is based on measurements in live UMTS networks in Vienna, Austria and the results are compared to the ob- tained statistics of measurements in a reference network where the transmission of data was performed without interference from other cells, multipath propagation or fading effects. Following the obtained statistics we develop a two-state model which is capable of describ- ing the measured statistics properly. Furthermore, we analyse the correlation properties of the channel and we show a two-layer model for an improved representation of the statistical dependencies between the errors of the UMTS DL DCH. I. I NTRODUCTION In digital wireless transmission systems the link errors have a great impact on the performance of the higher layer protocols [1] due to the statistical dependencies between the errors. In contrast to wired links, in wireless systems the errors are often grouped together in bursts because of the high variability of the radio link. That influences the functionality of the higher layer protocols more severely than the equally distributed transmission errors of wired sys- tems. Thus, it is all-important to be aware of the error characteristics of the used channel when developing new algorithms for multimedia signal processing or optimising the protocol stack for wireless transmission systems. Due to the high flexibility - especially of the 3G mobile communication networks - lots of different and new applications and protocols, as well as multimedia signal processing algorithms or cross-layer optimisa- tion methods like in [2] are currently developed for UMTS. The development procedure highly depends on the properties of the underlying system. Therefore, it is very important to analyse the error characteristics of the UMTS DCH (Dedicated Channel) beforehand as it is used by most of the services. In several publications the behaviour of the UMTS DCH has been investigated by link level or physical layer simulations [3]. Our approach in this paper is to provide an analysis of the error characteristics of the UMTS DCH based on measurements in live UMTS networks in Vienna, Austria. We are particularly fo- cussing on the error characteristics of the UMTS DCH in the DL (Down Link) with a throughput of 384kbit/s for static scenarios. This is the typical configuration when using a UMTS PCMCIA datacard or a mobile phone as a modem in connection with a notebook. For evaluation of the performance of protocols or algorithms, exhaustive simulations have to be per- formed where the lower layers are represented e.g. by stochastic models. In this paper we develop models, following the measured statistics, which are capable of describing the error characteristics of the UMTS DL DCH for 384kbit/s properly for the static case. In order to observe the influence of the propagation effects like multipath propagation or fading 1 on the error characteristics, we are comparing the measure- ment results of the live network with the results of the reference network. In the reference network the measurements can be performed with the mobile enclosed within a shielding box which is directly connected to the antenna connector of the Node B. This document is organised as follows. In Sec- tion II the measurement setup will be explained and the utilised parameters for traffic and system will be presented. Section III shows the analysis of the error characteristics of the UMTS DL DCH. First an analysis of the gap- and burstlengths 2 will be presented and then we show the results of the analysis of the correlation properties. In Section IV one possible way of modelling the error characteristics via a two- state model is presented. Furthermore, a two-layer model as an enhancement for the two-state model is shown, which is capable of describing the correlation properties more accurately. Finally, in Section V a summary and conclusions will be given. 1 Fading is present in a small amount also in static scenarios. 2 Gaplength/burstlength is the number of subsequently received error free/erroneous transport blocks. 8th International Symposium on DSP and Communication Systems, DSPCS'2005" & "4th Workshop on the Internet, Telecommunications and Signal Processing, WITSP'2005", Noosa Heads (Sunshine Coast, Australia), 19-21 December 2005
7
Embed
Measurement based Analysis and Modelling of … based Analysis and Modelling of UMTS DCH Error Characteristics for Static Scenarios Wolfgang Karner, Markus Rupp Institute of Communications
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Measurement based Analysis and Modellingof UMTS DCH Error Characteristics for
Static ScenariosWolfgang Karner, Markus Rupp
Institute of Communications and Radio-Frequency EngineeringVienna University of Technology, Austria
Abstract— For the evaluation of higher layer proto-cols, the simulation of cross-layer optimisation methodsor the development of new algorithms for multimediasignal processing it is essential to know the exact errorcharacteristics of the underlying channels. In this paperwe show the analysis of the error characteristics of theUMTS DL (Down Link) DCH (Dedicated Channel) with384kbit/s especially for static scenarios. The analysisis based on measurements in live UMTS networks inVienna, Austria and the results are compared to the ob-tained statistics of measurements in a reference networkwhere the transmission of data was performed withoutinterference from other cells, multipath propagationor fading effects. Following the obtained statistics wedevelop a two-state model which is capable of describ-ing the measured statistics properly. Furthermore, weanalyse the correlation properties of the channel and weshow a two-layer model for an improved representationof the statistical dependencies between the errors of theUMTS DL DCH.
I. INTRODUCTION
In digital wireless transmission systems the linkerrors have a great impact on the performance ofthe higher layer protocols [1] due to the statisticaldependencies between the errors. In contrast to wiredlinks, in wireless systems the errors are often groupedtogether in bursts because of the high variability ofthe radio link. That influences the functionality ofthe higher layer protocols more severely than theequally distributed transmission errors of wired sys-tems. Thus, it is all-important to be aware of the errorcharacteristics of the used channel when developingnew algorithms for multimedia signal processing oroptimising the protocol stack for wireless transmissionsystems.
Due to the high flexibility - especially of the 3Gmobile communication networks - lots of different andnew applications and protocols, as well as multimediasignal processing algorithms or cross-layer optimisa-tion methods like in [2] are currently developed forUMTS. The development procedure highly dependson the properties of the underlying system. Therefore,it is very important to analyse the error characteristicsof the UMTS DCH (Dedicated Channel) beforehandas it is used by most of the services.
In several publications the behaviour of the UMTSDCH has been investigated by link level or physicallayer simulations [3]. Our approach in this paper is toprovide an analysis of the error characteristics of theUMTS DCH based on measurements in live UMTSnetworks in Vienna, Austria. We are particularly fo-cussing on the error characteristics of the UMTS DCHin the DL (Down Link) with a throughput of 384kbit/sfor static scenarios. This is the typical configurationwhen using a UMTS PCMCIA datacard or a mobilephone as a modem in connection with a notebook.
For evaluation of the performance of protocols oralgorithms, exhaustive simulations have to be per-formed where the lower layers are represented e.g. bystochastic models. In this paper we develop models,following the measured statistics, which are capableof describing the error characteristics of the UMTSDL DCH for 384kbit/s properly for the static case.
In order to observe the influence of the propagationeffects like multipath propagation or fading1 on theerror characteristics, we are comparing the measure-ment results of the live network with the resultsof the reference network. In the reference networkthe measurements can be performed with the mobileenclosed within a shielding box which is directlyconnected to the antenna connector of the Node B.
This document is organised as follows. In Sec-tion II the measurement setup will be explained andthe utilised parameters for traffic and system willbe presented. Section III shows the analysis of theerror characteristics of the UMTS DL DCH. First ananalysis of the gap- and burstlengths2 will be presentedand then we show the results of the analysis ofthe correlation properties. In Section IV one possibleway of modelling the error characteristics via a two-state model is presented. Furthermore, a two-layermodel as an enhancement for the two-state model isshown, which is capable of describing the correlationproperties more accurately. Finally, in Section V asummary and conclusions will be given.
1Fading is present in a small amount also in static scenarios.2Gaplength/burstlength is the number of subsequently received
error free/erroneous transport blocks.
8th International Symposium on DSP and Communication Systems, DSPCS'2005"& "4th Workshop on the Internet, Telecommunications and Signal Processing,
WITSP'2005", Noosa Heads (Sunshine Coast, Australia), 19-21 December 2005
II. MEASUREMENT SETUP
A. General Setup
The measurements results for this work are basedon measurements in three live UMTS networks ofthree different operators in the city center of Vi-enna, Austria. Measurements in the three networkshave shown great similarity for the static case [4].Therefore, we are presenting in this document themeasurements out of only one operator’s live network,but at three different locations. Additionally, we showthe measurements performed in the reference network(a separate network for acceptance testing) of thesame operator, where corresponding parameters canbe adjusted.
USB
PC
RNC SGSN GGSN
Node B
TU-ViennaInternet
UuIub Iu/PS Gn Gi
UDP const. kbit/s
TEMS Investigation
WCDMA
Iub Iu/PS
RNC
Node B
PC
SGSNGGSN
TU-ViennaInternet
USB
UDP const. kbit/s
TEMS Investigation
WCDMA
Gn Gi
shielding
box
Fig. 1. Measurement setup, live network/reference network.
For the measurements a UDP data stream with a bitrate of 360kbit/s (372kbit/s incl. UDP/IP overhead)in DL was used. The data was sent from a PClocated at the Institute of Communications and RadioFrequency Engineering at the Vienna University ofTechnology to a notebook using a UMTS terminal as amodem. As depicted in Fig. 1 (upper part), during themeasurements in the live network the UDP data streamgoes from the PC over University LAN (ethernet),internet, UMTS core network and over the UMTS airinterface (with a 384kbit/s radio access bearer assignedduring the measurements in DL) to a UMTS mobilewhich is connected via USB (Universal Serial Bus) tothe notebook. In order to represent a static scenariothe mobile terminal was lying on a table in an officeor living room with only few movements of objects orpersons around the mobile during the measurements.
In case of the measurements in the reference net-work the signal is routed via coaxial cable to ashielding box directly from the antenna connector ofthe Node B as can be seen in Fig. 1 (lower part). In thatshielding box the UMTS terminal is enclosed in orderto avoid inter-cell interference, multipath propagationeffects and fading.
The statistics of CPICH Ec/I0 (Common PilotChannel chip energy to noise and interference ratio)
location 1
0
2000
4000
6000
8000
10000
12000
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
CPICH Ec/I0 [dB]
his
tog
ram
location 2
0
10000
20000
30000
40000
50000
60000
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
CPICH Ec/I0 [dB]
his
tog
ram
location 3
0
2000
4000
6000
8000
10000
12000
14000
16000
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
CPICH Ec/I0 [dB]
his
tog
ram
reference network
0
5000
10000
15000
20000
25000
30000
35000
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
CPICH Ec/I0 [dB]
his
tog
ram
Fig. 2. CPICH Ec/I0 at diff. locations in the live network and inthe reference network.
location 1
0
1000
2000
3000
4000
5000
6000
7000
-100 -95 -90 -85 -80 -75 -70 -65 -60
CPICH RSCP [dBm]
his
tog
ram
location 2
0
5000
10000
15000
20000
25000
-100 -95 -90 -85 -80 -75 -70 -65 -60
CPICH RSCP [dBm]
his
tog
ram
location 3
0
1000
2000
3000
4000
5000
6000
7000
8000
-100 -95 -90 -85 -80 -75 -70 -65 -60
CPICH RSCP [dBm]
his
tog
ram
reference network
0
5000
10000
15000
20000
25000
30000
-100 -95 -90 -85 -80 -75 -70 -65 -60
CPICH RSCP [dBm]
his
tog
ram
Fig. 3. CPICH RSCP at diff. locations in the live network andin the reference network.
and CPICH RSCP (CPICH Received Signal CodePower) during the measurements are shown in Fig. 2and Fig. 3 respectively. There can be observed thatthe values for CPICH Ec/I0 and CPICH RSCP arein the same regions in the measurements of live-and reference network, but there is no fading in thereference network. There was additional traffic duringthe measurements in the live network, but the usedcells were not heavy loaded and there was a 384kbit/sbearer assigned for the measurement data stream inDL for the whole measurement period. Furthermore,measurements in heavy loaded cells have shown thatthe cell load has no impact on the DCH error charac-teristics.
A WCDMA TEMS mobile3 (Motorola A835) wasused as terminal. On the notebook the measurementsof the mobile have been captured by ‘TEMS Inves-tigation WCDMA 2.4’ software also by Ericsson. Byparsing the export files of this software tool, the BLER(Block Error Ratio), burstlengths, gaplengths and otherparameters have been analysed as explained further onin this paper.
Of course we are aware of the disadvantage ofmeasuring with one type of mobile terminal only.However, the Motorola A835 was the only TEMS mo-bile station available to us and capable of presentingthe CRC (Cyclic Redundancy Check) information ofthe received transport blocks. The error characteristicsof the Motorola A835 are assumed to be comparableto the error characteristics of other mobiles becausesimilar results have been observed when comparingthe measured statistics of the BLER (Block Error Ra-tio) of the Motorola A835 with the statistics collectedwith Sony Ericsson Z1010 mobile terminals.
B. Relevant UTRAN Parameters
In the considered UMTS networks in Vienna, therelevant system parameters are as follows.
As TrCH (Transport Channel) in the DL a DCH(Dedicated CHannel) and RLC AM (AcknowledgedMode) has been used. In addition, turbo coding and atransport block size of 336 bits has been selected. Forthe 384kbit/s bearer the SF (Spreading Factor) was 8and the TTI (Transmission Timing Interval) was 10mswith 12 transport blocks per TTI.
Another very important parameter of the UTRANfor evaluating the error characteristics of the DCHis the BLER quality target value for the outer loopTPC (Transmit Power Control) which was set to 1%in the networks used for the measurements. As aconsequence the closed outer loop TPC tries to adjustthe SIR (Signal to Interference Ratio) target for theclosed inner loop (fast) TPC in a way that the requiredBLER quality (1% in our case) is satisfied.
III. ANALYSIS OF DL DCH ERRORCHARACTERISTICS
A. Analysis of gaplength and burstlength
In [4] we have shown a method for analysing theerror characteristics of the UMTS DCH based on TTIs,which means we have been looking at subsequent TTIsand if they contain an erroneous transport block ornot. The reason for using that method of analysingthe DCH was that in scenarios with mobility theprobability is 80-85% to receive all 12 transport blockserroneous within one TTI - considering only erroneousTTIs.
In static scenarios, the probability of receiving alltransport blocks (trbks) of a TTI in error (12 trbks in
3A modified Motorola A835 for TEMS data logging, offered byEricsson [5]. Modifications are only done in order to provide internalmeasurements via the interface.
Fig. 4. Measured number of erroneous transport blocks per TTI.
case of the 384kbit/s bearer), is only around 20% ascan be seen in Fig. 4. The probability of having onlyone erroneous transport block within one erroneousTTI is as well around 20% in the measurements of thelive network and even more than 70% in the referencenetwork.
Due to that fact, the analysis of the error character-istics of the UMTS DCH for the static case should bebased on transport blocks rather than on TTIs.
Fig. 5. Analysis of DCH error characteristics based on transportblocks.
An analysis method for the error characteristics ofthe DCH by the measured data is depicted in Fig. 5.The method builds on the observation of the state ofthe transport blocks. We are looking particularly forthe number of subsequent error free trbks what wecall the gaplength while the number of subsequenterroneous trbks is called burstlength. The definitionof an error burst is according to [6] where a burst is agroup of bits in which two successive erroneous bitsare always separated by less than a given number Lof correct bits. In our case L is equal to zero.
The statistics of gaplengths and burstlengths ob-tained from measurements in the static case - atdifferent locations in the live network and in thereference network - is presented in Figs. 6 and 7.
As can be seen in Fig. 6, there are two areas wherethe gaplengths occur with a higher probability. One isthe part of long gaps with more than 1000 subsequenterror free trbks and the other part is with gaps ≤ 12.Our assumption for the reason of the long gaps isthe period of the OLPC (Outer Loop Power Control)when it is built like proposed in [7]. In the idealcase (without any disturbance) the period of the OLPCwould be 100 for reaching a BLER target of 1% andthe gaplength would be 99. Due to the burstiness ofthe errors the period (and therefore the gaplengths)
Fig. 6. Measured gaplength in number of trbks, diff. locations.
become > 1000. The short gaps could be attributed tothe multipath propagation and fading because they areless probable in the reference network. Another reasonfor having gaps ≤ 12 could be due to effects of turbocoding. As shown in [8], the turbo coder produces biterrors which are equally distributed across the codingblock (TTI). We can see in Fig. 8 that the erroneoustransport blocks have also almost equally distributedpositions within the TTI in the live network. Fig. 9illustrates that short gaps can result in such a case.
Fig. 7. Measured burstlength in number of trbks, diff. locations.
Fig. 7 shows the measured burstlengths (numberof subsequent erroneous transport blocks) depicted asempirical cdf. It can be observed that the bursts havea length between one and 25 and the probability ofburstlength one is almost 90% in the reference networkand around 50% in the live network. At burstlengthsof 12 and 24, larger steps occur in the empirical cdfwhich is due to the fact that also in the static casethe probability of receiving all (12) transport blocksin error within one TTI is quite high.
B. Analysis of DL DCH Error Correlation Properties
From the statistics of gap- and burstlengths(Figs. 6, 7) we can already come to the conclusionthat there is correlation between the error states ofsubsequent transport blocks due to the burstiness ofthe errors. As can be observed, the errors are groupedtogether in bursts of lengths ≤ 24 trbks, separated bygaps with > 1000 trbks.
1 2 3 4 5 6 7 8 9 10 11 120
0.02
0.04
0.06
0.08
0.1
0.12
Positions of err. trbks within TTI
P(p
ositi
ons
in T
TI)
location 1
1 2 3 4 5 6 7 8 9 10 11 120
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Positions of err. trbks within TTI
P(p
ositi
ons
in T
TI)
location 2
1 2 3 4 5 6 7 8 9 10 11 120
0.02
0.04
0.06
0.08
0.1
Positions of err. trbks within TTI
P(p
ositi
ons
in T
TI)
location 3
1 2 3 4 5 6 7 8 9 10 11 120
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Positions of err. trbks within TTI
P(p
ositi
ons
in T
TI)
reference network
Fig. 8. Probability of positions of erroneous transport blocks withinTTI.
TTI, 10ms
trbk
Fig. 9. Schematic illustration of erroneous transport blocks withinone TTI.
Another very important property is the statistical de-pendency between subsequent gaps, subsequent burstsand between gaps and bursts. In Fig. 10 the schematicillustration of uncorrelated (memoryless) and corre-lated gaps and bursts is shown. When looking at theautocorrelation function of subsequent gaps in Fig. 11,one can observe the present correlation between atleast the nearest neighbours of subsequent gaps. Toanalyse the statistical independency between two suc-cessive gaps, the simple and conditional probabilities(conditioned on the length of the previous gap) of
burstlength gaplength
burstlength: iid
gaplength: iid
uncorrelated (memoryless):
iid iid
… erroneous trbk
… error free trbk
correlated:stat. dependency
stat. dependency
stat. dependency
stat. dependency
Fig. 10. Dependencies between subsequent gaps and bursts.
having a gaplength of ≤ 12 and > 12§, are shown inFig. 12. As the simple and conditional probabilities aredifferent, there is no statistical independence betweensubsequent gaplengths. Similar results can be shownfor subsequent burstlengths and between a gap and thefollowing burst. In order to capture these statisticaldependencies appropriately we propose the two-layermodel presented in Section IV B.
Fig. 12. Conditional probabilities of subsequent gaplengths.
IV. MODELLING OF DL DCH ERRORCHARACTERISTICS
In literature, many different ways of modelling theerror sequences of digital mobile channels have been
§The probabilities of ≤ 12 and > 12 have been chosen forthat analysis in view of the following modelling of the errorcharacteristics in Section IV.
proposed beginning with [9], [10] and [11] by the useof Markov chains, hidden Markov models [12] [13] orseveral other methods. In order to describe the desiredstatistics of gaplengths and burstlengths ideally, an N -state Markov model with large N would be necessary.As our intention is to keep the complexity of themodel as small as possible our first approach is atwo-state model (renewal process), which only hasa few easy determinable parameters and is suitablefor modeling the error characteristics of the UMTSDCH with uncorrelated gaps and burst. This modellingapproach is explained in the following subsection.
A. Two-state Model
The goal in modelling the error characteristics ofthe UMTS DCH in this subsection is to generate asequence of transport blocks with the correct statisticsof gaplengths and burstlengths. As already mentionedin Section III (Fig.6), there are two main areas withhigh probability of gaplengths - long gaps with morethan 1000 subsequent error free trbks and short gapswith ≤ 12 trbks. We are using the same partitioning ofgaplengths for modelling of the statistics. As shown inFig. 13 for location 1 in the live network, the statisticsof small gaplengths can properly be modelled via aWeibull distribution with scale parameter as 1.2 andshape parameter be 0.7.
Fig. 14. Measurement vs. simulation of gaplengths > 12 trbks,location 1.
The two-parameter Weibull CDF is given by
F (x|a, b) =
∫x
0
ba−btb−1e−( t
a)b
dt (1)
where a and b are scale and shape parameters, re-spectively. A Weibull distribution also meets the sta-tistics of the gaps with > 12 trbks (Fig. 14) witha = 3350 and b = 2.0167, again shown for themeasurement at location 1 in the live network. Thecombination of these two Weibull distributions now iscapable of describing the total statistics of measuredgaplengths adequately (see Fig. 15). The comparisonbetween the measured and simulated statistics of theburstlengths at location 1 in the live network is shownin Fig. 16, where again a Weibull distribution is usedfor modelling, supplemented by additional steps at theburstlengths of 12 and 24.
The reason for using a Weibull distribution in mod-elling gap and burst statistics is the high flexibilityof the two-parameter Weibull distribution [14]. Byproper adjustment of the shape parameter, exponential-, Rayleigh-, normal- and even other distributions canbe met or approximated.
Fig. 15. Measurement vs. simulation of gaplengths (two-partWeibull distribution), location 1.
0 5 10 15 20 250
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
burstlength (number of transport blocks)
empi
rica
l CD
F
measurement, location 1
simulation, Weibull (1.1, 0.55) + peak at 12, 24
Fig. 16. Measurement vs. simulation of burstlengths, location 1.
Following the explained method of describing ad-equate statistics for gap- and burstlengths we arriveat a two-state model as shown in Fig. 17, where inone state correct transport blocks and in the other stateerroneous transport blocks are generated. In the correctstate, the number of subsequent error free transport
blocks (gaplength) is calculated via a two-part Weibulldistributed random number and the number of sub-sequent erroneous transport blocks (burstlength) inthe error state is calculated via a simple Weibulldistributed random number but with additional peaksat burstlengths of 12 and 24. After each calculationof either burst- or gaplength, the state is changed.Therefore, for modelling the error characteristics ofthe UMTS DCH in this way, a total number of nineparameters are required. These are: four parameters forthe two Weibull distributions and one parameter forthe separation between the distributions in the correctstate. Additionally, two parameters for the Weibulldistribution and two parameters for determination ofthe peaks at 12 and 24 in the erroneous state.
state: trbks correct
gaplength = two-part Weibull
distributed random number
state: trbks in error
burstlength = Weibull distributed
random number + peaks at 12,24
Fig. 17. Two-state model.
B. Two-Layer Model for modelling Correlation Prop-erties
The two-state model mentioned in the previoussubsection is a renewable process, meaning there isno dependency between subsequent gaps, subsequentbursts or neither between a gap and the following burst(see the schematic illustration in Fig. 10). On the otherhand - as shown in Section III B - there is statisticaldependency between subsequent gaps, bursts and alsobetween a gap and the following burst. Therefore wepropose the following two-layer model (kind of hiddenMarkov model) where correlation between gaps andbursts is introduced via the upper layer two-stateMarkov chain (Fig. 18).
pss pllpsl
pls
state: trbks
correct
state: trbks
in error
state: trbks
correct
state: trbks
in error
short state
gaplength <= 12
long state
gaplength > 12
Fig. 18. Two-layer two-state Markov model.
The model has one ’short state’ (gaps with ≤ 12trbks) and one ’long state’ (gaps with > 12 trbks)and the corresponding transition probabilities are theconditional probabilities out of Fig. 12. After enteringa certain state, first the gaplength is calculated via aWeibull distributed random number for either the smallor the large gaplengths according to the separationin Section III and depending on the current upperlayer state. Then, while staying in the same state ofthe upper layer Markov chain, a burst is calculatedwith different probabilities of either creating a Weibulldistributed burstlength or a peak at 12 or 24, againdepending on the actual upper layer state. After that
the state of the upper layer Markov chain is changedand the procedure begins again with calculating thegaplength.
Thus, via the upper layer Markov chain, the modelintroduces correlation between subsequent gaps andalso between a gap and the following burst by us-ing 11 parameters. Two parameters are necessary fordetermination of the transition probabilities of theupper layer Markov chain. Another two parametersare used for the Weibull distributions of each of thecorrect states and only two parameters more for theWeibull distributions of both erroneous states as theWeibull distributions are the same for the erroneousstates of both upper layer states. What depends onthe current state of the upper layer Markov chainin case of the erroneous states is the probability ofgenerating Weibull distributed burstlenghts or havinga burst with lengths 12 or 24. For these probabilitiesand the separation between a burstlength of 12 and 24another three parameters are needed.
We can see the improvements in modelling the errorcharacteristics by using this correlated model in thefollowing subsection.
C. Simulation ResultsIn this subsection we are comparing the simulated
results of the two presented models with the mea-sured statistics. In Fig. 19 the empirical CDF ofthe error weight probability P(m,n), the probabilityof exactly m errors occurring in a block of n, isshown which we call the Block Error Ratio (BLER)with n equal to 2400 trbks. We can see that theuncorrelated two-state Model is able to describe themeasured statistics quite well although it does notmeet the correct correlation properties of the channel.This can be seen in the absence of the bigger stepsat 0.5% and 1%. The correlated model is bringingan observable enhancement to the problem by usingtwo more parameters. Of course, the two-layer modelonly catches a part of the correlation characteristicsof the channel. Therefore, only small improvementscompared to the uncorrelated model can be achieved.
0 0.5 1 1.5 20
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
BLER (%)
empi
rica
l CD
F
BLER measuredBLER simulated, uncorrelated modelBLER simulated, correlated model
Fig. 19. Comparison of BLER measurements and simulation resultsof the models, location 1.
V. SUMMARY AND CONCLUSIONS
We have shown the analysis of the error charac-teristics of the UMTS DCH (Dedicated Channel) inDL (Down Link) for static scenarios, based on mea-surements in live UMTS networks in Vienna, Austria.We present statistics of gap- and burstlengths whichare furthermore compared to the statistics obtainedin a reference network, where transmission withoutmultipath propagation or fading is possible. It is shownthat these statistics can all be perfectly matched byWeibull distributions because of the great flexibilitythe two-parameter Weibull distribution offers in itsshape parameter. Following the measured statistics wehave developed a two-state model describing a renew-able process which is capable of meeting the statisticsof gap- and burstlengths properly. Additionally weanalyse the correlation properties between subsequentgaps and bursts and it is shown that there is nostatistical independence. Thus, we further propose atwo-layer model (kind of hidden Markov model) asan enhancement to the simple two-state model.
ACKNOWLEDGEMENTS
We thank mobilkom austria AG&CoKG for tech-nical and financial support of this work. The viewsexpressed in this paper are those of the authors anddo not necessarily reflect the views within mobilkomaustria AG&CoKG.
REFERENCES
[1] M. Zorzi, R.R. Rao, “Perspectives on the Impact of ErrorStatistics on Protocols for Wireless Networks”, IEEE PersonalCommunications, Vol. 6, pp. 32–40, Oct. 1999.
[2] O. Nemethova, W. Karner, A. Al-Moghrabi, M. Rupp, “Cross-Layer Error Detection for H.264 Video over UMTS,” in Proc. ofWireless Personal Multimedia Communications 2005 (WPMC2005), Aalborg, Denmark, Sept. 2005.
[3] J.J. Olmos, S.Ruiz, “Transport Block Error Rates for UTRA-FDD Downlink with transmission Diversity and Turbo coding,”in Proc. IEEE 13th PIMRC 2002, vol.1, pp 31-35, Sept. 2002.
[4] W. Karner, P. Svoboda, M. Rupp, “A UMTS DL DCH ErrorModel Based on Measurements in Live Networks,” in Proc.of 12th International Conference on Telecommunications 2005(ICT 2005), Capetown, South Africa, May 2005.
[5] http://www.ericsson.com/products/hp/TEMS Products pa.shtml[6] ITU-T Rec. M.60, 3008; ITU-T Rec. Q.9, 0222.[7] H. Holma and A. Toskala, WCDMA for UMTS, Radio Access
For Third Generation Mobile Communications, Third Edition,John Wiley & Sons, Ltd., 2004.
[8] F. Navratil, “Fehlerkorrektur im physical Layer des UMTS,”Master’s thesis, Institute of Communications and Radio-Frequency Engineering, Vienna University of Technology, Aus-tria, Nov. 2001.
[9] E.N. Gilbert, “Capacity of a burst-noise channel,” Bell SystemsTechnical Journal, vol. 39, pp. 1253–1265, Sept. 1960.
[10] E.O. Elliot, “Estimates of error rates for codes on burst-noisechannels,” Bell Systems Technical Journal, vol. 42, pp. 1977–1997, Sept. 1963.