TECHNICAL REPORT. An NS Module for Simulation of HSDPA. Mats Folke Sara Landström. Luleå University of Technology

Transcription

1 TECHNICAL REPORT 2006:03 An NS Module for Simulation of HSDPA Mats Folke Sara Landström Luleå University of Technology Technical Report - Department of Computer Science and Electrical Engineering Division of Computer Communication 2006:03 - ISSN: ISRN: LTU-TR--06/03--SE

2 An NS Module for Simulation of HSDPA Sara Landström and Mats Folke Division of Systems and Interaction Department of Computer Science and Electrical Engineering Luleå University of Technology {sara.landstrom, April 2006

3

4 Contents 1 Introduction Simulation of wireless systems HSDPA Media Access Control Fast Link Adaptation Transport block selection process Fast Hybrid ARQ Fast Scheduling Radio Link Control Cell selection Admission control Propagation environment Path loss, shadowing, and multi-path fading Interference Block error rates Cell planning Mobility 13 5 Limitations and Future Work 14

5

6 1 Introduction We have implemented functionality of the High Speed Downlink Packet Access (HSDPA) mode as an extension to ns-2 (the Network Simulator version 2 [16]). Ns-2 is a well-known simulation platform frequently used in network research. It includes detailed models of transport protocols, queueing strategies, etc. The code is open source, which is in line with the research ideal that results and methods should be made available both for validation of results and for further studies. We see a need for our extension because mobile cellular systems are merging with the wired Internet creating an all-ip network. This surfaces problems related to TCP performance, Quality of Service, user behaviour to name a few. Modelling of high-speed cellular networks, from an Internet point of view, is therefore needed. When the current TCP/IP protocol suite was designed, assumptions were made about the delay and loss characteristics of the underlying links [13]. The initial wireless IP networks suffered from problems in situations when the assumptions were not met and numerous solutions have been proposed, see [3, 15, 19] for some examples. Our extension module is designed to study problems related to TCP performance when several TCP flows are set to compete for the resources of a shared radio channel with varying bit rates. It is therefore tted with detailed models of TCP and factors that will have a signicant impact on its performance, like radio-block scheduling for example. These factors appear on longer time scales. Effects on shorter time scales are not simulated, since the effect of those are dwarfed by the factors previously mentioned. This allows for savings in simulation time because simplied models of the effects on short time scales can be used. The purpose of this report is to describe the models used in our extension. For better understanding of our models, we also include descriptions of the system modelled. This report is not a manual for performing simulations with our extension. The extension itself and other documentation can be found at [4]. The work on this extension was started by Ulf Bodin, who used it to produce the results presented in [5]. It has since been constantly improved to produce results in [14, 10, 8, 9]. 1.1 Simulation of wireless systems The simulation of mobile cellular systems is a challenge due to the inherent complexity of the radio link modelling and all the functionality necessary to support transportation over the wireless media. There are two main areas of simulation of mobile cellular systems, link level simulations and system level simulations. Link level simulations study problems related to single link transmission, selecting appropriate modulation schemes and coding rates during certain conditions. A link level simulator often uses at the chip or symbol resolution. It therefore has to have a high resolution in time since the effects studied appear on very short time scales. With a link level simulator, the behaviour of users and network protocols appear on longer time scales and are therefore simplied, if modelled at all. A system level simulator on the other hand is designed to study problems related to longer time scales, such as performance of transport protocols like TCP and user-experienced quality, and multi-user interaction over the shared channel. It therefore includes detailed models of events appearing on longer time scales. Effects on 1

7 shorter time scales not affecting the research problems can therefore be modelled in a simplied manner. The separation into system level and link level simulations reduces simulator complexity, provides greater flexibility and allows for timesavings to be made. Our extension module simulates system level characteristics and makes use of link level results produced in [11]. The network simulator could simulate link layer behaviour, but its strength is the fairly detailed models of higher layer protocols. Since we are not interested in link layer performance, but instead problems associated with system performance, we will probably not extend the simulator to include more accurate link level models. Another extension of ns-2 that models HSDPA is Eurane (Enhanced UMTS Radio Access Network Extensions for ns-2) [25]. It differs from our extension in that it contains better support for internal transport protocols and channels including the high-speed downlink shared channel (HS-DSCH), but it lacks support for mobility and it can only simulate one cell. Eurane was developed within the SEACORN project [20]. Our extension module contains algorithms for estimating SIR, performing channel allocation, link adaption and handovers. The interference depends on the offered load, which is created by the user, application and transport layer behaviours. Furthermore, the number of UEs (user equipments) present in a cell and their locations, as well as mobility patterns are important factors. Models for path loss, fast and slow fading and power control are also included. The focus of this report is to describe these mechanisms, both in general terms and the chosen models. 2 HSDPA In Europe the European Telecommunications Standard Institute (ETSI) [7] unites manufacturers, network operators, service providers, administrators, research bodies and users with the common interest of producing telecommunications standards. ETSI is active within the 3rd Generation Partnership Project (3GPP) [1], which was established in December 1998 and brings together a number of telecommunications standards bodies. 3GPP attend to the development of globally applicable technical specications and technical reports for a 3rd Generation Mobile System based on evolved GSM technology. The ETSI 3G standard is called the Universal Mobile Telecommunications System (UMTS). Within in the International Telecommunication Union (ITU) [12] the corresponding standard is referred to as the International Mobile Telecommunications-2000 (IMT-2000) and covers a global perspective. ITU is an international organisation within the United Nations System [26] where governments and the private sector coordinate global telecom networks and services. WCDMA, as specied by 3GPP, is becoming the leading IMT-2000 standard [18]. WCDMA stands for Wideband Code Division Multiple Access and is the radio access technology chosen by ETSI and the corresponding Japanese agency [2]. Release 5 of the 3GPP specications satisfy the IMT-2000 requirements and development has been taken beyond 3G towards 3.5G, when it comes to datarates and delay, by the introduction of High-Speed Downlink Packet Access (HSDPA) [21]. HSDPA is an extension to UMTS, in the same way as the Enhanced Data rates for GSM Evolution (EDGE) [22] is an extension to GSM and may in theory provide data rates of up to 14 Mbps. HSDPA was designed for bursty trafc, such as the trafc carried by TCP 2

8 in the Internet, i.e. web trafc and . Such trafc consists of short periods of high trafc followed by periods of silence. Shared channels, such as HS-DSCH, are targeted at bursty packet data, where resources would be underutilised if they were reserved for a single user. Voice trafc is not carried by TCP and can be viewed as constant bitrate trafc with on and off periods, due to silence suppression. In addition to the increased data rates, HSDPA may achieve shorter delay and a jitter reduction compared to previous technologies. There are basically three enabling techniques, fast link adaptation, fast hybrid ARQ and fast scheduling[18]. These techniques all rely on a rapid adaptation of the transmission parameters to the instantaneous channel conditions. In particular, a new transport channel, the High Speed Downlink Shared Channel (HS-DSCH), to be used for best-effort packet data in HSDPA, puts this into practice. In the UTRAN (Universal Terrestrial Radio Access Network) architecture there are several radio network controllers (RNCs) that bridge to the core network. The RNCs control groups of Node Bs (base stations). User Equipments (UEs) connect to the network through these base stations. Previously scheduling, transport format selection and retransmissions have been handled by the RNC, but this functionality has now been moved to a new Node B instance, called the Media Access Control for HSDPA (MAC-hs) as shown in Figure 1. The MAC unit and the Radio Link Control (RLC) protocol together form the radio link layer. Our module focuses on the MAC-hs layer in the Node B. We therefore do not model the RNC nor the core network. We believe that the impact on higher layers from those components is small compared to the impact from the MAC-hs layer. The reason for this is that the core network, including the RNC act as a routing network without bottlenecks, which can be simplied to a single link. A situation where detailed modelling of the core network would be of interest is during a hand-over scenario, where the propagation delay of the core network may differ. This difference is however negligible compared to other delays such as scheduling delay or possible delay when performing the hand-over. 2.1 Media Access Control Radio resources assigned to HSDPA in the form of base station power and codes are distributed by the MAC-hs unit. HSDPA has an almost constant power for shorter timescales, but for HS-DSCH the power can be controlled on the level of individual TTIs. This keeps the interference level down. The total power assigned to HSDPA can however vary due to power demands from other downlink channels. These variations are believed to be rather slow, hence we have not implemented any such long-term power variation. The short-term power control of HS-DSCH is implemented. In HSDPA, a pool of common channelisation codes with a spreading factor of 16 is assigned to HS-DSCH. Code multiplexing of the 15 channelisation codes is possible for up to four users in order to increase the spectral utilisation and this is supported in our implementation. It is further assumed that a xed power and a constant number of channel codes have been reserved for HS-DSCH. Due to the interaction between transport layer mechanisms and resource allocation algorithms, resource allocation necessarily has to be part of the system simulation. The primary reason for the 3

9 Core Network RNC RLC MAC Node B MAC hs PHY Node B MAC hs PHY UE UE UE UE UE Figure 1: Modied UTRAN architecture strong dependency is that the resource allocation algorithm affects the round trip times, which is a factor in TCP's congestion detection mechanisms. As a consequence, TCP might reduce its sending rate and there may not be any blocks to schedule. Varying round trip times also halts the acknowledgement clock of TCP, which also is negative to TCP performance. Because of these problems we have included fairly detailed models of resource allocation, such as transport block selection (Section 2.1.2), radio-block scheduling (Section 2.1.4), admission and admission control (Section 2.3.1) Fast Link Adaptation Link adaptation in HSDPA is the ability to adapt the modulation type and coding rate to the quality of the radio link, while keeping a xed spreading factor of 16. The interference variations can be reduced when adaptive modulation and coding (AMC) is used instead of power adaptation to compensate for link quality changes. Two modulation techniques have been chosen for HS- DSCH, QPSK and 16QAM. QPSK is used as a complement to the higher order 16QAM scheme, which has a higher spectral efciency in terms of bits/s/hz but is less robust to interference. By allocating a large part of the downlink power to a single user at a time, better signal- 4

10 to-interference conditions are possible and hence higher-order modulation can be applied. In a system with AMC, users close to Node B are typically assigned higher order modulation with higher code rates, and the modulation-order and/or code rate generally decreases as the distance to Node B increases. Channel coding protects digital data from errors by selectively introducing redundancies in the transmitted data. Our module does not explicitly model modulation nor coding rates. Instead we have a precalculated table that translates the estimated signal-to-interference ratio, the number of channelisation codes left and the amount of data available of scheduling to a transport block size. Explicit modelling of modulation and coding rates would be too time consuming, regarding simulation run time, to justify such an implementation. This is because the effects of a more precise implementation would appear on too short time-scales to impact TCP performance. More detailed modelling is therefore not necessary because the varying bit rate as a result of varying radio conditions has the dominant influence on TCP performance Transport block selection process The coding and modulation schemes can be seen in Table The values in the table are based on the work presented by Hosein in [11], where the following equation was presented: R(t) = (0.2 x(t) 3 + B) log 2 ( x(t)/10 ) (1) where R(t) is the bitrate in bits per second, x(t) is the signal to interference ratio in db and B is the effective bandwidth in Hz. In [11] Hosein used this equation as an approximation of an existing modulation and coding scheme (MCS) mapping. In [23] methods for determining transport block size can be found in section Based on measurements from the UE a value called k t is computed which used to lookup the transport block size in a table. Using the TB sizes found in the table of Annex A in [23] and Equation 1 we can estimate at which level the SIR has to be in order for a particular block to be selected. Table contains these combinations. When assigning a transport block our algorithm will take the available data, power and channelisation codes and SIR target into consideration. Thus, it makes sure to not use more power than necessary, keeping the interference level low. Power control directly influences SIR and thereby also dynamic channel allocation, hand-over decisions and vice versa. Early work [5, 10, 14] done with this simulator used another mapping based on values taken from [18], which was presented in those papers. Our new table increases the granularity in the bit rates, which closer models the real system. The bit rates of the new table can more closely map to the bit rates of the services, thus reducing overhead. This is particularly useful when simulating low-bit-rate services such as voice over IP (VoIP) Fast Hybrid ARQ Errors in channel quality prediction can be compensated for using fast Hybrid Automatic Repeat request (HARQ). HARQ combines forward error coding using the UMTS turbo code with a N-channel Stop and Wait protocol (SAW). The SAW protocol sends a packet, then waits for the acknowledgement in order to determine whether the next packet should be a retransmission 5

11 k t Modulation Codes TB-size (bits) Bit rate (kbps) SIR (db) 63 QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QPSK QAM QAM QAM QAM QAM QAM QAM QAM Table 1: This table is used to map the estimated SIR-value to a transport block size. corresponding bit-rates, channelisation codes and modulation are also shown. The 6

12 or a new packet. To increase the efciency of the SAW protocol, there are N channels on which data is alternately sent. While one is inactive waiting for an acknowledgement, data can be sent on another channel. The number of channels are commonly set to six, which results in a retransmission delay of 12 ms. One such channel is often referred to as a HARQ process. There are several different strategies used to choose which data to send when a retransmission is requested. In Chase Combining the exact same data is retransmitted. Additional decoding gain can be achieved by letting the retransmission consist partly, called Partial Incremental Redundancy (IR), or entirely (Full IR) of new parity bits. Maximum decoding gain has to be traded versus buffer requirements for soft combining [6]. The decoder at the receiver combines these multiple versions of the transmitted packets weighted by the received signal-to-noise ratio (SNR). In our model no coding gain is assumed, that is decoding of the data after a retransmission is as likely to succeed as the operation was after the rst transmission. The retransmission techniques could be simulated in a link level simulator yielding different error rates based on the number of times a packet has been transmitted, but the delay caused by the retransmissions has to be added at the system level. In our simulator the queue of the sender of the damaged radio-block is blocked for 12 ms, which corresponds to the delay of six HARQ-processes. After this delay the user competes for resources like any other user. We believe that the delay is the component which will affect the behaviour of upper layers the most, hence we do not do any more detailed modelling of the HARQ. This simplied model also ensures in-order delivery of packets and requires no need for complex priority queues. A somewhat more correct model could for example be to lower the bit error rate of the retransmitted radio block at the expense of bit rate. Such extension would however limited impact on TCP performance. It is therefore questionable if more detailed modelling of HARQ can be justied Fast Scheduling HS-DSCH is primarily shared in the time domain. Different users can be scheduled every transmission time interval (TTI) which lasts for 2 ms. The high granularity makes it possible to quickly adapt to the instantaneous channel conditions. All the codes and power available for HS-DSCH can be allocated to one user for a TTI. To ll up the payload up to four simultaneous transmissions can be allowed. A scheduler can base its decisions on channel quality, terminal capability, QoS class and power/code availability. Often a tradeoff between fairness and spectrum utilisation must be made by the scheduling algorithm. There are currently seven scheduling algorithms implemented. Since we have the possibility of sending to four users in a timeslot, the scheduler ranks up to four users according to its scheduling criteria. Thereafter the algorithm for code multiplexing decides on how many of these four users to which data will be sent to. The goal of the code multiplexing is to maximise the system throughput with respect to both power and channelisation codes. It will not reorder the users ranked by the scheduler. In the equations describing the schedulers the following terminology is used. The momentary maximum bit rate achievable by user i with respect to both radio conditions and available data is denoted by r i. The average bit rate of user i is denoted by r i and can be calculated from the beginning of a data flow or using an EWMA lter. For some schedulers it is interesting to talk 7

13 about a minimum bit rate which is denoted by r min. The time spent by the foremost packet in the queue of user i is denoted by d i. The time a packet may spend in the queue before it is prioritised is denoted by d prio. The momentary signal-to-interference ratio (SIR) of user i is denoted by S i. The average SIR of user i is denoted by S i. The user which is picked by the scheduler is denoted by i. The Round Robin (RR) scheduler lets the users take turns to transmit in an orderly fashion. The actual throughput may differ from user to user since coding and modulation is chosen according to the prediction of the channel conditions for that particular user. A secondary slot user also consumes its turn and has to wait for the next round, as well as users that do not have anything to send when it is their turn. Starvation is not possible with an RR scheduler, since a user eventually will get data through. It is therefore believed that the RR scheduler is some what fair in user throughput. The Max-SIR scheduler will pick the user for which i = arg max i {S i }. (2) This scheduler aims at obtaining high spectrum utilisation. With SIR scheduling it is possible to starve a user who has low SIR. The SIR scheduler may select users with good radio conditions, but with very little data. This separates this scheduler from the Max-rate scheduler, which takes the amount of data available into consideration. The Max-rate (MR) scheduler which will optimise for system throughput by selecting the user for i = arg max i {r i }. (3) As the Max-SIR scheduler, this scheduler may also starve some users. However, because of the high throughput there will be a large turn-around among the users, i.e. users will nish their flows quickly and exit the system. The problem of starvation is less imminent with this scheduler compared to the Max-SIR scheduler. The Proportional Fair (PF) scheduler will pick the user for which i = arg max i { ri r i }. (4) This scheduler is more fair than Max-rate because it considers the average bit rate r i in addition to r i. During periods when a user is not scheduled r i will decrease. A lower r i makes the user more likely for selection. Due to this property the PF scheduler will not suffer from starvation. 8

14 The Relative SIR scheduler is a variant of Max-SIR, the same way the PF scheduler is a variant of the Max-rate scheduler. It will pick the user for which i = arg max i { Si S i }. (5) This means that a user which currently has good radio conditions in relation to its long-term history, will be picked over a user which has even better radio conditions, but has had so for quite some time. In this way, the scheduler will be more fair than Max-SIR while trying to maintain a high spectrum utilisation. The MR scheduler with a minimum bitrate requirement (MRmin) picks the user for which i = arg max{r i (1 + βe β(r i r min ) )}. (6) i This scheduler supports multiple trafc classes, in that r min can be tuned for different trafc classes. We have used this scheduler to differentiate between VoIP trafc and web trafc [9]. The scheduler will promote users having an average throughput (r i ) below the minimum (r min ). The parameter β determines at what rate this promotion increases and it can be set according to trafc class. The MR scheduler with a delay requirement (MRdelay) is especially useful for trafc with a delay bound such as VoIP or streaming. It picks the user for which arg max{r i } if d i < d prio, i = i (7) arg max{d i + R max } otherwise i R max is the maximum theoretical bitrate possible. By adding this to d i, we make the scheduler prioritise the flows where the foremost packet in the queue has been delayed longer than d prio. This scheduler has freedom to maximise system throughput while still reserving some capacity for flows which are delay-sensitive. 2.2 Radio Link Control The most essential Radio Link Control (RLC) service is packet management. The RLC takes Service Data Units (SDUs) from the upper layer and transforms them into Packet Data Units (PDUs) through segmentation, concatenation and padding (if there is no data left). After transmission the SDUs are reassembled and may be delivered in preserved order to the upper layer. In acknowledged data transfer mode faulty PDUs may be retransmitted. The RLC functionality has not been explicitly implemented. Instead the downlink data are directly divided into radio blocks whose sizes are determined by the radio conditions, channelisation codes and available payload space in the current TTI. The headers added by the RLC layer are included in the implementation. Data may come from several ongoing sessions. 9

15 An explicit implementation of the RLC functionality could include more layers of segmentation and reassembly. This would be transparent to TCP and thus have no effect on simulation results. Such addition would merely increase the simulation run time. 2.3 Cell selection The user equipment itself should decide which is the best cell for it to receive HSDPA transmissions from and inform the network. This procedure is called Fast Cell Selection. Decision metrics apart from the radio propagation conditions may be power and code availability. The implemented cell selection algorithm is executed at the start of each scheduling interval (TTI). It picks the cell where the highest gain in Watt is possible and whose admission control admits the user. In reality there are several practical complications. For example the UEs' requests are only guaranteed to be received correctly by one basestation, therefore the requested basestation may not be reached. State information has to be transferred between basestations and there are probably limitations to how often UEs can move between cells etc. These problems are related to hand-overs. This has not been the main area of study. A more detailed implementation of hand-over would add more complexity, but would not affect the performance of aggregated TCP flows in such an extent that an implementation would be justied. If the performance of individual TCP flows or if problems directly related to hand-overs were to be studied, a more detailed model of hand-overs should be implemented Admission control Admission control is an important part in any QoS system. HSDPA contains mechanisms and frameworks for admission control on HS-DSCH, but there are no standardised algorithms or methods. Instead the manufacturers and operators are developing methods of their own. Our implementation contains three algorithms for admission control. Once admitted a user cannot be dropped by the system, but it can deregister itself upon completion of a transfer. When beginning a new transfer the user must then be readmitted. The rst algorithm admits all users which is equivalent to no admission control. With this method some cells may be heavily overloaded causing deteriorating quality. This method is only used as a reference case for other more realistic ones. The second algorithm has a maximum number of users which are allowed to be admitted to a cell. This method will admit a new user if the current number of admitted users is less than the maximum number of allowed users. The maximum number is statically congured before the simulation starts. The third algorithm admits users based on throughput and is the most advanced one. It will admit a user if the current mean throughput among the already admitted users in the cell is greater than some minimum bit rate set in advance. This can viewed as a simplied algorithm providing some minimum bit rate guarantee. 10

16 3 Propagation environment In order to compute interference and other channel metrics a system perspective is required. In reality channel parameters are measured on the actual signals traversing the system, but in our case we use models to predict the values of the parameters and delays to model the feedback loops. Status reports for the HARQ protocol, measurement reports for the link adaptation process and fast cell selection information travelling uplink are therefore not included. Signal propagation in a mobile radio environment is affected by three independent phenomena: Deterministic path loss with distance. Random slow shadowing. Random fast multipath fading. In this section the models for these mechanisms will be accounted for, as well as the interference considerations that are made. 3.1 Path loss, shadowing, and multi-path fading Large scale propagation models are used to predict the signal strength, when the transmitter and the receiver are far apart. It is generally assumed that the average received signal power decreases logarithmically with distance in both outdoor and indoor environments. This simple model does not consider the surrounding environment. Measurements indicate that the path loss at a particular location is random and normally distributed in db about the mean value. Equation 8 combines the expressions for path loss and shadowing. ( ) d P L(d)[dB] = P L(d 0 ) + 10n log + X σ (8) d 0 The path loss exponent, n, sets the rate at which the path loss increases with distance. For urban cellular radio n normally ranges from 2.7 to 3.5. The surroundings are accounted for by the factor X σ which is a zero-mean Gaussian distributed random variable in db with standard deviation σ. The distance between the transmitter and the receiver is d and d 0 is a close-in reference distance which is determined from measurements close to the transmitter. Our module implements Equation 8 to model the path-loss and shadow fading. There is a correlation between the surroundings at the user's new position and the old. Therefore the new shadowing value is a weighted sum of the old shadowing value and a new value drawn from the Gaussian distribution. The degree of correlation between the values depends on how much the user has moved since the last estimate and the correlation distance. The weights are constant for stationary users but become variable when users move. Typically, the autocorrelation prole is rst order negative exponential as shown in Figure 2. The correlation distance is usually chosen between metres, which corresponds to the average size of the obstructions, commonly buildings. 11

17 1 Autocorrelation profile for shadowing Autocorrelation /e 0.3 Correlation Distance 40 metres Distance moved between updates of the shadowing values [metres] Figure 2: Typical autocorrelation prole for shadowing, for a correlation distance of 40 metres. Relative to the shadowing effect, the randomly delayed, reflected, scattered, and diffracted signal components cause rapid fluctuations in the received signal strength. We model this behaviour using a Typical Urban model [24]. The multi-path fading is dependent on the speed of the users, which is further described in Section Interference WCDMA systems are generally interference limited when the frequency reuse of the system is 1, i.e., the entire frequency spectra is used in every cell. There is no hard limit for the number of users possible, instead performance decreases gradually with increasing load. In WCDMA/FDD interference between the uplink and downlink is completely avoided by a duplex separation of 190 MHz. For both the uplink and the downlink, there are sources of interference in the own cell and in co-channel cells in the area. The inter-cell interference is limited only by the distance between the transmitter and the receiver (UE or base station), which is attempting to receive another signal. Signal attenuation is discussed in Section 3.1. Self interference is interference caused by the desired signal itself through multipath propagation or from replicas of the signal arriving from different base stations. Parts of the signal can be reflected off buildings and mountains, such that it appears to be a signal from a different sender when it arrives. Self interference is modelled by a xed percent of the signal strength. It is possible to make this constant value time-varying, which would lead to a more complex model, perhaps closer to the real world. However, if the number of users is large, it is likely that the power variations due to self interference will be averaged out thus making it valid to use a constant. 12

18 HS-DSCH has a number of codes to be shared primarily in the time domain, but code multiplexing is possible. In a multipath propagation environment there is interactions between transmissions when data is sent in parallell to more than one user, since the orthogonality of the codes is not preserved. As with self interference the simplest model is to consider a constant part of the power that is used for other sessions as contributing to the interference level. If there is no multipath propagation the system is fully orthogonal and the orthogonality factor α is equal to 1. In multipath channels the value is between Block error rates In our model, the block error rate (BLER) is dependent on the radio conditions db is the rst value in Table on Page 6 and if the radio conditions is worse than that, we model this with increased BLER of 50%. The reason is that no modulation coding scheme (MCS) is regarded to be suitable for such a low SIR and it is likely that the information sent will be received in error. For users with a SIR higher or equal to db the BLER is 20%. All block errors are randomly uniformly distributed. 20% is a relatively high number, but it compensates for the fact that the channel estimation is perfect. The SIR used to choose MCS is the true SIR, no errors in the link quality estimation has yet been implemented. A more exact mapping to the block error probabilities for a given SIR and MCS is possible. The effects thereof would however appear on too short time-scales to have impact on the performance of TCP. 3.4 Cell planning Our cell plan contains seven sites each having three sectorised cells, in total 21 cells. The border effect comes from the fact that the outermost cells will have an incomplete number of neighbours, compared to the inner cells. This makes the interference calculations incorrect. To mitigate the border effect we can add more cells, making more cells have a complete number of neighbouring cells. However, no matter how many tiers of cells are added, the outermost tier will always have an incomplete set of neighboring cells. Adding more cells can be inefcient since more cells have to be simulated. In our module we use a technique called ``wrap-around''. The seven simulated sites also become ``shadow sites'' in the second tier to give all the rst tier cells a full set of neighbours. This method is known as ``wrap-around'', since the interference disappearing from the cells wraps around the cell plan and reappears on the other side. This technique is also applied to users moving off the cell plan. 4 Mobility One of the main motivations for wireless access is mobility, but to what extent are the users moving? The implemented module supports user mobility. A user that drops off the edge of the 1 In the ITU Vehicular A channel α = 0.6 and in the ITU Pedestrian A channel α =

19 simulated cell plan will reappear at the edge of another cell. Start positions, speeds and directions are all part of the mobility model used. The mobility model recommended in [17] suggests that the UE speeds should be drawn from the distribution given in Table 2 and kept for a simulation run. This speed distribution is implemented. By introducing mobility a more dynamic system can be modelled, but one disadvantage with user mobility is the possibility for several users ending up in one (or a few cells), causing a skewness between the cells regarding the load. One solution to this is to use the speed for input to path-loss and fast-fading calculations, but in fact keeping the mobile stationary. Speed (kph) Percentage Table 2: Speed distribution 5 Limitations and Future Work The primary applications intended for HSDPA are web browsing and video on demand. Both these services generate application data trafc only on the downlink, but the transport layer and application layer protocols may require acknowledgements to be sent on the uplink. The simulator is built to study the downlink trafc. Hence, uplink trafc is simplied. The uplink trafc consists mostly of acknowledgements generated by the transport protocols. This link can be simulated using existing ns-2 models, i.e., by a xed delay and possibly a certain packet drop rate. The implicit assumption is thus that the system will not be interference limited in the uplink, but rather in the downlink. Although uplink functionality is implemented, the initial simulations have been made less complex by simulating a lossless uplink with a xed delay [5]. Current trends suggest that voice over IP (VoIP) may also be run over HSDPA. This means that the amount of trafc flowing in the up- and downlink would be almost the same. A more precise modelling of the uplink may therefore be necessary to implement. 14

20 References [1] 3rd Generation Partnership Project (3GPP) [2] Association of Radio Industries and Businesses (ARIB). english/, [3] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, and R. H. Katz. A comparison of mechanisms for improving TCP performance over wireless links. IEEE/ACM Transactions on Networking, 5(6):756769, [4] U. Bodin, M. Folke, and S. Landström. HS-DSCH extension to ns-2. luth.se/csee/csn/mobivin/wikka/hssim, [5] U. Bodin and A. Simonsson. Effects on TCP from Radio-Block Scheduling in WCDMA High Speed Downlink Shared Channel. In QoFIS, pages , [6] M. Döttling, J. Michel, and B. Raaf. Hybrid ARQ and Adaptive Modulation and Coding Schemes for High Speed Downlink Packet Access. In PIMRC, [7] European Telecommunications Standard Institute (ETSI) [8] M. Folke and U. Bodin. On the Influence of User Behaviour and Admission Control on System Performance in HS-DSCH. In Proceedings of IEEE Vehicular Technology Conference, Melbourne, Australia, May [9] M. Folke, S. Landström, U. Bodin, and S. Wänstedt. Scheduling Support for Mixed Conversational and Background Trafc over HSDPA. Submitted for review, March [10] M. Folke, S. Landström, and U. Bodin. On the TCP Minimum Retransmission Timeout in a High-speed Cellular Network. In Proceedings of the Eleventh European Wireless Conference, Nicosia, Cyprus, Apr [11] P. A. Hosein. Capacity of packetized voice services over time-shared wireless packet data channels. In INFOCOM 2005, volume 3, pages , March [12] International Mobile Telecommunications [13] V. Jacobson. Congestion avoidance and control. In Proceedings of ACM SIGCOMM Conference, pages , [14] S. Landström, L. Åke Larzon, and U. Bodin. Congestion control in a high speed radio environment. In International Conference on Wireless Networks, Las Vegas, Jun [15] R. Ludwig and R. H. Katz. ``The Eifel algorithm: making TCP robust against spurious retransmissions''. ACM SIGCOMM Computer Communication Review, 30(1):3033,

21 [16] S. McCanne and S. Floyd. The Network Simulator ns-2. nsnam/ns, [17] Motorola. Evaluation Methods for High Speed Downlink Packet Accesss (HSDPA). Technical report, 3GPP, Jul [18] S. Parkvall, E. Dahlman, P. Frenger, P. Beming, and M. Persson. The Evolution of WCDMA Towards Higher Speed Downlink Packet Data Access. In Proceedings of IEEE Vehicular Technology Conference, pages , May [19] M. Patel, N. Tanna, P. Patel, and R. Naerjee. TCP over Wireless Networks: Issues, Challanges and Survey of Solutions. Technical report, Computer Science Department, The University of Texas at Dallas, [20] SEACORN web page [21] A. Sultan et al. Overview of 3GPP Release 5. Technical report, ETSI Mobile Competence Centre, September [22] A. Sultan et al. Overview of 3GPP Release 99. Technical Report Version xx/07/04, ETSI Mobile Competence Centre, [23] Technical Specication Group Radio Access Network (TSG-RAN). Medium Access Control (MAC) protocol specication (Release 6). Technical Report TS V6.6.0, 3rd Generation Partnership Project (3GPP), September [24] Technical Specication Group Radio and Network (TSG-RAN). Deployment aspects. Technical Report TR V6.0.0, 3rd Generation Partnership Project (3GPP), December [25] Twente Institute for Wireless and Mobile Communications. Eurane web page. http: // [26] United Nations System of Organizations