Hybrid model for throughput evaluation of orthogonal frequency division multiple access networks

Data throughput is an important metric used in the performance evaluation of the next generation cellular networks such as longterm evolution (LTE) and LTE-advanced. To evaluate the performance of these networks, Monte Carlo simulation schemes are usually used. Such simulations do not provide the throughput of intermediate call state; instead it gives the overall performance of this network. The authors propose a hybrid model consisting of both analysis and simulation. The benefit of this model is that the throughput of any possible call state in the system can be evaluated. Here, the probability of possible call distribution is first obtained by analysis, which is used as input to the event-driven-based simulator to calculate the throughput of a call state. Comparison is made between throughput obtained from the author’s hybrid model with that obtained from event-driven-based simulation. Numerical results are presented and show good agreement between both the proposed hybrid model and the simulation. The maximum difference of relative throughput between their hybrid model and the simulation is found in the interval of (0.04 and 1.06%) over a range of call arrival rates, mean holding times and number of resource blocks in the system.


Introduction
With increasing demand for mobile data services, orthogonal frequency division multiplexing (OFDM) has become one of the most promising radio interface technologies for future-generation wireless networks.It has already been adopted by systems such as long-term evolution (LTE) and LTE-advanced (LTE-A) [1].The goal of LTE is to improve the spectral efficiency and hence increase the network capacity, improve services and lower costs.In an orthogonal frequency division multiple access (OFDMA) network, transmission is achieved by transmitting data via multiple orthogonal channels.The system allocates power and transmission rate adaptively and optimally among the subcarriers to achieve high data throughput.Owing to the use of multiple orthogonal channels, OFDM also performs equalisation and is consequently robust to inter-symbol interference and frequency-selective fading [2,3].
One of the most important aspects of any commercial mobile network deployment is its information carrying capacity, and data throughput is one fundamental parameter in the capacity planning for cellular system deployment.LTE-A is being standardised, so performance evaluation is essential in order to provide insights into competing contributions prior to deployment.System-level performance studies of emerging broadband wireless networks such as LTE-A is typically simulation-based.Such simulations are oriented towards assessing the base station (BS) performance, and do not consider user performance.To estimate user performance, an appropriate performance metric that reflects the throughput is needed.
Ismail and Matalgah [4] have evaluated the performance of code division multiple access (CDMA) network by simulation approach, whereas Kelif and Alman [5] and Kostas and Lee [6] have evaluated the performance of CDMA and time division multiple access (TDMA) network by analytical approach, respectively.Ahn and Wang [7] have evaluated the performance of the OFDMA network based on analytic and simulation approach separately.However, none of the papers mentioned above have considered the combined approach of analysis and simulation.On the other hand, Wu and Sakurai [8] have considered a hybrid simulation/analysis approach, where a detailed rate distribution obtained via simulation is used as input to a generalised processor sharing queue model, but does not consider the user-level.To the best of the authors' knowledge, none of the papers have considered a hybrid model to evaluate user performance.There may be the case where we need to know the throughput of any particular call state from the user point of view.Simulation approach do not provide the throughput of any intermediate call state because it is not possible to obtain this data, whereas the analytical approach provides the average performance.The benefit of the hybrid model is that the throughput of any possible call state can be evaluated.
In this paper, we present a novel hybrid model of analysis/simulation for determining the average data throughput of a system from the user point of view, where a detailed probability of call distribution is obtained from analytical expressions and used as input in the simulator to evaluate the throughput performance.Such a model has the benefit of evaluating the performance of any specific call state from the user perspective.
The rest of the paper is organised as follows.Section 2 describes the system considered for evaluation.Section 3 describes the analytical model for calculating the probability of call distribution in the system, simulation model for evaluation and our proposed hybrid model.The simulation configuration is described in Section 4 and then Section 5 presents and compares results obtained from both conventional simulation and our hybrid approach.Finally, Section 7 provides conclusions.

System description
For OFDMA-based networks, user data are divided and modulated onto a large number of narrow-band subcarriers in the frequency domain, and each of them is modulated by low rate data [9].The subcarriers are orthogonal to each other, meaning that cross-talk between the subcarriers is eliminated and inter-carrier guard bands are not required.The orthogonality among the subcarriers prevents inter-subcarrier interference because the subcarrier's spectrum has nulls located at the centre frequencies of adjacent subcarriers [9].A group of consecutive subcarriers is known as a subchannel.Moreover, the time domain is split into consecutive frames that are in turn divided into time slots called OFDM symbols.As a multiple access technique, OFDMA offers the possibility of enhancing the spectral efficiency of networks by assigning distinct OFDM symbols or subchannels to distinct users, thus taking advantage of their diverse time and frequency channel conditions as compared with TDMA and FDMA techniques.
For LTE, downlink transmission is based on OFDMA.The radio resources can be considered as a frequency-time resource grid as illustrated in Fig. 1.In the frequency domain, the radio spectrum is divided into a number of narrow subcarriers of 15 kHz (in addition to 15 kHz subcarrier spacing, a reduced subcarrier spacing of 7.5 kHz with twice OFDM symbol time is also defined for LTE which targets multicast-broadcast single-frequency networkbased multicast/broadcast transmissions [10].)In the time domain, a frame of 10 ms duration is divided into ten subframes of 1 ms each.Each subframe is further divided into two time slots of 0.5 ms each.Each time slot then consists of six or seven OFDM symbols depending on the length of cyclic prefix (normal or extended cyclic prefix) [9].A grid of one subcarrier (15 kHz) in the frequency domain and one OFDM symbol (0.5 ms) in the time domain is known as one resource element, whereas a grid of 12 adjacent subcarriers (12 × 15 = 180 kHz) and one OFDM symbol (0.5 ms) is known as one resource block (RB).Hence, an RB is a rectangular block of resource elements, which spans 12 adjacent subcarriers in the frequency domain and 7 OFDM symbols in the time domain (180 kHz × 0.5 ms).In LTE, an RB is also known as a 'subchannel', and from now on we refer to an RB as a subchannel.Depending on the transmission bandwidth, a downlink carrier comprises a variable number of subchannels in the frequency domain.The minimum bandwidth of 1.4 MHz corresponds to six RBs, whereas the maximum one of 20 MHz corresponds to 110 RB.The assignment of subchannels to users is carried out by the medium access control scheduler, and it is performed on a subframe-by-subframe basis, that is, each 1 ms.The scheduler decides which users are allowed to transmit on which subchannel.It should be noted that the minimum resource scheduling unit [From now on, when we refer to an RB, we refer to this minimum scheduling unit of two consecutive RBs, spanning 1 ms.] that the scheduler can assign to a user is comprised of two consecutive RBs and thus spans an entire subframe.

Model description
Consider cellular layout as shown in Fig. 2, where each hexagon is divided into three cells.Each cell is equipped with one transmit antenna and each user equipment (UE) has one receive antenna.Users arrive in a cell according to a Poisson process and are uniformly distributed within the cell.

Notations
For the sake of clarity, let us introduce some general notations.Let † J be the total number of cells in the system.† U be the number of user/subscriber class types, that is, voice, video, streaming etc.For simplicity, we will use call service for all types of service.† M j be the maximum number of users in cell j.† R j be the number of subchannels in cell j(1 ≤ j ≤ J ). † λ u, j be the mean arrival rate of type-u (1 ≤ u ≤ U ) class user in cell j.† λ nu,j be the mean arrival rate of type-u class new user in cell j.† λ hu,j be the mean arrival rate of type-u class handoff user in cell j.† h u, j be the mean holding time of type-u class user in cell j.† r u, j be the cell residence time of type-u class user in cell j and is exponentially distributed with mean 1/r u, j .† μ u, j be the mean service rate of type-u class user in cell j.† p ju, ku be the probability that user of type-u class moves from cell j to a neighbouring cell k, given that it moves to a neighbouring cell before the call is completed such that J k=1 p ju, ku = 1.† X u, j (t) denotes a random variable related to the number of type-u class users in progress in cell j at time t.

Steady-state distribution of calls
We assume that new users are generated in cell j according to a Poisson random process with mean arrival rate λ nu,j and the requested call connection time is exponentially distributed with mean μ nu,j = 1/t u, j [11].Both the inter-arrival time (1/λ u, j ) and the mean holding time (h u, j ) are exponentially independent and identically distributed random variables [11][12][13].In a cellular network, the handoff traffic is considered more important than a new arriving traffic, because the forced termination of an ongoing call is considered less desirable than the blocking of a new call [14][15][16].The priority schemes of how to handle the handoff traffic depend on network designers.One of the popular scheme is 'guard channel scheme' [11,14,16,17], where a fixed number of subchannels in a given cell is reserved for handoff traffic.For example, in a cell with R subchannels, r g number of subchannels is reserved for handoff traffic.We assume that the network topology does not change before the steady state is reached.The state of a cell j at time t can be written as The state of the network at time t is with state space where 0 and n j is a vector representing the number of calls in cell j.The statistically stationary distribution of calls in the system is given by [11,18] where G j is the normalisation constant chosen such that the sum of the probabilities of all possible call states in any cell j is 1, and the sum of the product of probabilities of all possible call configurations in the system is 1.That is and The normalisation constant is written as [11] where in which 1[l < R j − r g ] is the indicator function taking value 1 if the statement l < R j − r g is true, else zero; λ nu, j is the arrival rate of new user of type-u class and λ hu,j is the arrival rate of handoff user of type-u class and is given by Chao and Li [11] l hu, j = J k=1 where p ku, ju is the probability that user u moves from a cell k to cell j.Based on the above equations, we can compute the probability of cell j, π j (n j ), being in a particular state of call and the probability of the system, π(n), of a specific state n = (n 1 , . . ., n j , . . ., n J ) [ S of the system, as illustrated in an example in Appendix.
Since the constant term G j is a function of R j , we shall write it as G j (R j ), that is where for convenience g u, j is defined as The marginal probability that there are n u, j class u calls in the cell j is [11] for n u, j = 1, …, R j , where G (u) j (R j − n u, j ) is the normalisation constant of cell j with class u calls removed, that is

Blocking probability
According to the guard channel scheme, a new call in a cell j gains a subchannel if it finds that there are less than R j − r g calls in the cell and that there is at least one subchannel available, otherwise, the new call is blocked in cell j and will be cleared from the system.On the other hand, a handoff call into a cell j gains a subchannel if it finds at least one subchannel available, otherwise, it is blocked.
(1) New call blocking probability: a class u new call is accepted to cell j with probability [11] R j −r g −1 Then, the blocking probability for a class u call in cell j is given as [11] where P u, j (B N ) represents the blocking probability of class u new call in cell j.Case 1: no handover and no reserved channel: suppose users move within the serving cell (i.e.no inter-cell mobility) and there is no policy of reserved guard band channels in order to maximise the spectral efficiency.Then, the probability that the user of type-u class moves from cell j to a neighbouring cell k is p ju, ku = 0 and the guard band channel, r g = 0.In this case, (8) reduces to λ u, j = λ nu, j , which is the arrival rate of new calls.In this case, the blocking probability of class u new calls can be evaluated as For user of voice type service, the blocking probability of new call in a cell can be evaluated as (2) Handoff call blocking probability: a class u handoff call is accepted to cell j with probability [11] v n v, j ≤R j −1 Then, the blocking probability for a class u call to cell j is given as [11] where P u, j (B H ) represents the blocking probability of class u handoff call to cell j.

Channel modelling
The medium between the transmitting and the receiving antennas is known as the 'channel'.The characteristics of radio signal changes as it travels from the transmitter antenna to the receiver antenna.The characteristics depend on the parameters such as distance between these two antennas, propagation scenario (e.g.outdoor-to-outdoor, outdoor-to-indoor, indoor-to-indoor etc.) and the surrounding environment (e.g.buildings, trees etc.).The received signal can be estimated if we have a suitable model of the medium.This model of the medium is called the 'channel model'.The radio channel propagation is typically modelled as the combination of three main effects: the mean path loss, the shadowing generally characterised as log-normal [19,20] and the fading typically modelled as Rayleigh [21].In OFDMA system, the data are multiplexed over a large number of narrow-band subcarriers that are spaced apart at separate frequencies; the subchannel consists of parallel, flat and non-frequency-selective fading.The received signal is then only impacted by slow fading.
(1) Path loss model: path loss is the distance dependent mean attenuation of signal as it propagates through space.A suitable model of path loss depends on the parameters such as type of the environment (e.g.macrocell, microcell, indoor etc.), the propagation medium (e.g.outdoor-to-outdoor, outdoor-to-indoor, indoor-to-indoor etc.), the carrier frequency and the distance.The path loss model recommended by 3rd generation partnership project (3GPP) [1] for outdoor macrocells at a carrier frequency of 2 GHz is modelled as (L u, j ) dB = 128.1 + 37.6 log 10 (d u, j ) where L u, j is the path loss in dB from cell j for user u and d u, j is the distance in km from the cell j to the user u.
(2) Auto-correlation shadow fading model: in reality, clutters from objects such as buildings, trees, terrain conditions etc. along the path of signal propagation differs for every path, and consequently signal attenuation varies from path to path.Shadow fading is used to model variations in the path loss because of such obstacles between the mobile and the BS.Shadow fading is also known as 'slow fading' [22].The effect of shadowing is commonly approximated by a log-normal distribution [22,23].Accordingly, the shadow fading in the path between a BS and a UE can be priori modelled using a log-normal random variable, L sh = N (m, s), where μ and σ are the mean and the standard deviation in dB, respectively.However, the modelling of shadow fading when considering the change of user's position is more intricate because of spatial auto-correlation between paths.The shadow fading process is auto-correlated in space, meaning that a moving UE may see similar shadow fading attenuations from the same BS at different but nearby locations.
A widely adopted auto-correlation model for shadow fading is the Gudmundson model [22], which defines the auto-correlation coefficient as follows where d cor is the decorrelation distance (which is defined as the distance at which the correlation coefficient ρ a falls to 0.5 [24]), and Δx is the distance between two positions.The auto-correlation of shadow fading can be implemented as follows.If L 1 sh is the log-normal component, N (m, s), in dB at position P 1 and L 2 sh is the log-normal component in dB at position P 2 , which is Δx away from P 1 , then L 2 sh can be modelled as a normally distributed random variable, L sh , in dB with mean m ′ and standard deviation s ′ as [24] L sh = L 2 sh = N m ′ , s ′ (23a)

Signal-to-interference plus noise ratio
For simplicity, uniform and equal transmission power is distributed on each subcarrier.Assuming that all subcarriers within a subchannel experience the same channel condition, the downlink signal-to-interference plus noise ratio (SINR) at the UE, u, on subchannel r connected to the cell j is given by g u, r, j = P u, r, j G u, r, j /(L u, j L sh ) i[C, i=j P u, r, j G u, r, j /(L u, j L sh )a r, i + N where P u, r, j and P u, r, i are the transmit power from the cells j and i, respectively; G u, r, j and G u, r, i are the antenna gains from the cells j and i, respectively; L u, j and L u, i are the path loss from the cells j and i, respectively; L sh is the attenuation because of shadowing; N is the thermal noise power; and α r, i is the subchannel allocation indicator, which is given by

Radio access bearer (RAB) efficiency
The RAB is the entity responsible for transporting radio frames of an application over the radio access of the network.From the estimated SINR, a suitable modulation and coding scheme (MCS) is selected for each user provided that the SINR satisfies the threshold for the selected MCS.The higher the SINR, the higher-order MCS is used satisfying the SINR threshold value.The RAB efficiency (defined in 3GPP 36.213Table 7.2.3-1[25]) is shown in Table 1.
In general, the RAB efficiency of a user on subchannel r is estimated as [27] where ξ r is the RAB efficiency (bits/symbol) on the subchannel r for the selected MCS, CR k is the coding rate of the MCS and M k is the number of constellation points of the MCS k , where k ∈ {1, 2, …, 15} represents a particular MCS as shown in Table 1.

User throughput
Once an MCS is selected, the bit-rate of the user u over the subchannel r can be estimated as [28] BR u, r = j r n r symbol r t where BR u, r is the bit-rate (bits/s), n r is the number of subcarriers in the subchannel r, symbol r is the number of data symbols in the subchannel r in t duration of a subframe.
Once the bit-rate is known, the throughput of the user u connected to the cell j over the subchannel r can be estimated as [28] T u, r, j = BR u, r 1 − 1 g u, r, j , MCS k (28) where ε(γ u, r, j , MCS k ) represents the block-error rate suffered by the user u over the subchannel r connected to the sector j, which is a function of its both SINR, γ u, r, j , and RAB MCS k .

Hybrid model
This model combines the analytic and simulation approach, which, unlike traditional simulation approaches, enables us to estimate throughput of any intermediate call state.The probability of call distribution being in a particular state in the system is obtained from analytic expressions and is used as input to the simulation to calculate the throughput of a cell.According to hybrid model, the average throughput for cell j is expressed as follows where n [ S is the possible call configurations in the system, represents the number of users in the cell j, π j (n j ) represents the probability that there are n j calls in cell j and T u, r, j is the simulation throughput (sim.thr.) of a user u on subchannel r in cell j obtained for a particular state of call in the system.The average throughput for the system is expressed as follows where π(n) represents the probability of state space n in the system, and T is the average system throughput obtained from simulation for a particular call configuration in the system.

Simulation configuration
To evaluate the performance of the system and validate our proposed hybrid model, an event-driven dynamic system-level simulation was used.The traffic is modelled by a homogeneous Poisson random process in each cell.To account for the dynamic behaviour of the incoming traffic pattern or service time and hence thereby obtain various traffic loads in the network, we have implemented two approaches.In the first approach, the mean holding time of users is fixed and the inter-arrival time is varied in order to model the dynamic behaviour of incoming traffic, whereas in the second approach, the inter-arrival time is fixed and the mean holding time is varied in order to model the dynamic service time.During the simulation, an event occurs when: (a) a user arrives and accesses a subchannel to connect to the network (b) the user moves position randomly (c) the user leaves the network and the subchannel is freed (d) the system triggers to log the simulator status indicator The logged data of the users such as SINR and throughput are obtained on a regular basis.The wireless channel for a user from a BS is selected randomly from available subchannels and remains the same as long as the user stays in the network.
A different level of adaptive MCS is selected from Table 1 when mapping the user's SINR to its achievable data throughput.The parameters in the simulation are consistent with the LTE downlink, and are listed in Table 2.
The process for evaluating the performance of the system in the hybrid model is described as follows: 1. Calculate the probability of the system being in a particular call configuration from the statistically stationary distribution of calls using (3)- (7). 2. Run the event-driven simulation for this particular configuration only and change the user position during the simulation (thus accounting for the effect of different locations when estimating the average throughput).3. Multiply the simulated throughput for this particular configuration obtained from the simulation with the probability calculated analytically for the system in a specific configuration.4. Repeat the simulation for all possible configurations of calls in the system. 5. Sum the throughput for all possible configurations for the overall system throughput.

Results
To validate the simulation, we have tested the results of probability of call distribution and probability of call blocking obtained from the simulation with that obtained by analysis for different traffic intensities.Fig. 3 shows that both the probability of call distribution for the highest possible call state and the probability of call blocking obtained by simulation are in-line with that obtained by analysis for different traffic intensities and different number of users.The performance of the system throughput obtained by our hybrid model was evaluated for different scenarios of traffic behaviour by changing the mean arrival rate, the mean holding time and the number of subchannels.

Varying arrival rate
In this case, the performance of the system was evaluated for different call arrival rates for different number of subchannels to account for the variation of inter-arrival traffic, whereas the mean holding time remained fixed.Table 3 and Fig. 4 show the performance of the average system throughput for different arrival rates.It is noted that the system throughput performance by the hybrid model is similar to the simulation for all arrival traffic patterns.The mean of the throughput error between the two methods is found to be −0.88% for R = 2, −7.60% for R = 3, −0.36% for R = 4 and −0.2% for R = 5.

Varying holding time
In this case, the performance of the system was evaluated for different mean holding times for different number of subchannels to account for the variation of service time, whereas the mean arrival rate remained fixed.Table 4 and Fig. 5 show the performance of the average system throughput for a range of mean holding times.It is noted that the system throughput performance by the hybrid model is similar to the simulation for all arrival traffic patterns.The mean of the throughput error between the two methods is found to be 0.94% for R = 2, −1.06% for R = 3, −0.67% for R = 4 and −0.66% for R = 5.

Conclusions
We have proposed a hybrid model consisting of analysis and simulation for the evaluation of average system data throughput compared with event-driven-based simulations.Our approach allows throughput of intermediate call state to be evaluated as well as the overall network throughput.To evaluate the performance by the hybrid model, a detailed probability of call distribution in the system is first obtained from analytical expressions of a statistically stationary distribution, which are used as an input to the simulator to calculate the system throughput.We compared the results of the hybrid model with those obtained from simulation.We tested the model for different parameters of user arrival rate, their mean holding time and different numbers of radio subchannels in the network.It has been found that the results of the hybrid model are in-line with the simulation-based results.The maximum difference of mean throughput error performance between the hybrid model and the simulation is found to be in the interval of (0.04 and 1.06%) for different call arrival rates, mean holding times and number of subchannels in the system.It was noted that for a large number of cells and users, the number of possible call configurations in the system is very large.In such a large possible number of call configurations, it is difficult to evaluate the system throughput by the hybrid model from the user point of view, because we need to evaluate the throughput for all possible call configurations.However, there may be the case where we need to know the throughput of any particular call configuration in the system from the user point of view.The simulation does not provide the throughput of any intermediate call state because it does not log the call state.The benefit of the hybrid model is that the throughput of any possible call state can be evaluated.

Table 3
System throughput performance with mean holding time, 1/μ = 3 min

Table 4
System throughput performance with mean arrival rate, λ = 2 calls/min