A Network Simulation Tool to Generate User Traffic and Analyze Quality of Experience for Hybrid Access Architecture Oscar D. Ramos-Cantor, Technische Universität Darmstadt, oscar.ramos@nt.tu-darmstadt.de, Germany Moritz Lossow, Telekom Innovation Laboratories, moritz.lossow@telekom.de, Germany Heinz Droste, Telekom Innovation Laboratories, heinz.droste@telekom.de, Germany Gerhard Kadel, Telekom Innovation Laboratories, gerhard.kadel@telekom.de, Germany Marius Pesavento, Technische Universität Darmstadt, mpesa@nt.tu-darmstadt.de, Germany Abstract A Hybrid Access Architecture, where users can be simultaneously served by different technologies, has been envisioned in order to increase the achievable data rates and enhance the user experience. Such an access can be provided to users where the costs of replacing existing technology are unmanageable and the complementary technology is underused. In order to understand the implications of a hybrid access between Digital Subscriber Line (DSL) and Long Term Evolution (LTE) for Downlink (DL), a Network Simulation Tool has been developed. The Traffic and QoE Simulator (TQoES) defines and models the type of services demanded by users, establishes an algorithm to select the kind of access to be used by specific services and simulates the behavior of users within an LTE network based on 3GPP recommendations. 1 Introduction With the advent of new high bandwidth services, such as HD streaming, cloud computing and IPTV, the global Internet traffic is being raising significantly in the latest years [1]. Depending on the employed technology, the users expectation is to have access to such high demanding services at anytime, anywhere. This is particularly true for fixed users connected to the Internet through a Digital Subscriber Line (DSL); in the case of mobile users, certain flexibility is still allowed as the users acceptance of the technology limitations is higher. In order to satisfy this expectation, deployment of broadband technologies such as Fiber to the Home (FTTH) and highly dense Long Term Evolution (LTE) network is an option. However, under certain scenarios the deployment costs of such technologies make the solution unfeasible; therefore, additional alternatives need to be explored. The Hybrid Access Architecture solution proposed by Deutsche Telekom AG [2], implies the simultaneous usage of two technologies, DSL and LTE in order to increase the maximum data rates provided to a user. In rural scenarios, deploying FTTH is expensive and the load of the LTE network is low; thus, the Hybrid Access is a promising solution to the increasing traffic demand. In this work a Matlab based Network Simulation Tool ( Traffic and QoE Simulator (TQoES)) is presented in order to study the implications of implementing the Hybrid Access Architecture in Suburban/Rural environments for downlink. The TQoES is based on four main modules. In the user generation module, the traffic services are associated to the user by following measured distributions of data volumes and busy probabilities in the busy hour. Home Gateway Core Network Internet Figure 1: Hybrid Access Architecture with DSL and LTE technologies The services are modeled by Two State Markov Processes and are defined in order to simulate typical IP services such as Voice over IP (VoIP), streaming and file downloading. The Hybrid Access module is in charge of splitting the traffic demand generated by the user into the DSL and LTE accesses; the criterion of using cheapest line first, i.e. DSL, is used, accessing the LTE network only when the demand is higher than the DSL capacity provided to the user. An LTE module is used to simulate the user behavior when accessing the wireless technology. The LTE module uses the Shannon bound provided by the 3rd Generation Partnership Project (3GPP) [3] in order to estimate the channel condition of the user; the scheduler used is a modified Round Robin scheduler to reduce complexity while keeping fairness with respect to assigned resources. Finally, a Quality of Experience (QoE) module is applied to evaluate the user experience per service type, where specific rules have been defined to assess whether the user is satisfied or not with the achievable data rates. The remainder of the paper is organized as follows. Section 2 introduces the Hybrid Access Architecture. Sections 3, 4 and 5 respectively describe the modeling approaches of the user, services and access. The model validation and exemplary results are examined in Sections
6 and 7, respectively. Finally, conclusions are given in Section 8. 2 Hybrid Access Architecture A diagram of the Hybrid Access Architecture is illustrated in Figure 1. The user selects services to be provided in the Downlink (DL); a Home Gateway applies the Hybrid Access Algorithm to distribute the traffic between the fixed (DSL) and wireless (LTE) connections. Each technology is connected via the Backhaul link to the provider s core network and then to the Internet. The TQoES emulates the behavior of the user in a second based, by demanding different services according to criteria to be explained in Section 4.3. A decision step is made in order to distribute the traffic between the two accesses. The cheapest line is used first, i.e. the LTE network is only used after the maximum DSL capacity is reached. If the user is accessing the LTE network, scheduling decisions are made based on the extended Round Robin scheduler explained in Section 5.2. Finally, transmission is assumed to be held without errors and the traffic from both accesses is added to assess the user experience per service type. 3 User Modeling In order to simulate the Hybrid Access, users need to be generated. A user is defined by a traffic demand composed of one or several traffic services, and a description of the accesses conditions, e.g. maximum throughput of the DSL connection and radio link quality (Signal to Interference plus Noise Ratio (SINR)) for the LTE access. The service selection is based on Cumulative Distribution Functions (CDFs) of the data volume and active time per user in the busy hour. These CDFss are reproduced by the TQoES and can be obtained from different sources. The input distributions used in the TQoES are based on anonymized measurements of four hour traffic from over 10000 DSL connections. In this paper, log normal distributed user demands and busy probabilities are selected as exemplary distributions, due to the confidentiality of the measurements. The parameters µ and σ of the distributions are provided in Table 1. The user demand distribution is for 4 hours and in bits. The log normal distribution s heavy tail is avoided by limiting the user demand to a maximum of 4 GB, which equals a cumulative distribution of 99.7 %. The lowest busy probability, which can be realized by the simulator is 0.1 %, i.e. 3.6 s in the busy hour. Therefore, busy probabilities lower than 0.1 % are not considered; this percentage is equivalent to a cumulative distribution of 8.8 10 12. Mean µ σ User demand 67.8 MB/h 20.7 1.3 Busy probability 14.2% -2.2 0.7 Table 1: Log normal distribution parameters 4 Service Modeling 4.1 Service types and QoE definition Three service types have been defined to be used by the TQoES with their respective QoE criteria. Each service type is thought to cover a group of Internet applications typically demanded by users, where the expectations vary from ensuring connectivity to reduce waiting times. The three services can be summarized as follows: Linear Real Time (LRT) Services, e.g. VoIP, have a low data rate requirement with respect to other kind of services. The user s expectation is to keep the call ongoing under satisfactory conditions to make the communication possible. In the TQoES, such a requirement is translated into the obligation to provide an instantaneous service throughput which is greater or equal to a minimum value. If the instantaneous service throughput is below the threshold the communication is not possible and the call would be dropped prematurely. Non Linear Real Time (NLRT) Services, e.g. streaming, require higher data rates than the LRT services. During streaming, the user expects an audio/video file to be played smoothly, i.e. without interruptions. For simulation, the QoE criteria used for NLRT services allows variable data rates through time (assuming buffering is taking place) but requires the average data rate to be greater or equal than a threshold. If the average service throughput is below the minimum value, it is interpreted that the file suffered too many interruptions, i.e. providing a bad user experience. Non Real Time (NRT) Services, e.g. file downloading, have the highest data rates and the loosest constraint from all service types. For the simulation it is enough to calculate and compare the download data rates. The highest it is, the better the user experience. 4.2 Service Model The traffic services described in section 4.1 are modeled by a Markov Process [4] as illustrated in Figure 2, where P a and P i are the active and idle probabilities respectively; and t aa, t ai, t ia and t ii are the transition probabilities describing the transition matrix T as [ ] taa t T = ai, (1) t ia t ii where the first subindex indicates the initial state and the second, the final state. The busy probability is obtained from the active time distribution used in the simulator. Thus, the idle probability is obtained as P i = 1 P a. (2)
t aa t ai Start P a P i t ii Data vol. & active time distributions User s data vol. (D Um ) & active time (P Um ) calculation t ia Figure 2: Markov Model representation for traffic services. In the TQoES, at each simulation step (snapshot) corresponding to 1s, a new state of the Markov Process is calculated by a stochastic process, taking into account the previous state and the transition matrix T. To calculate T, Equation 3 is used based on the Perron-Frobenius theorem [5], where λ is the second eigenvalue of the matrix T. λ can take values between ( 1 : 1) and is restricted further in order to keep the number of iterations required to stabilize the Markov Model, low. Service mix calculation (combinatorial) Data volume calculation for service mix (D Uc ) D Um D Uc? No t ia = 1 λ 1 + P i /P a, t aa = λ + t ia, t ai = 1 t ia, t ii = 1 t ia. (3) End Yes Figure 3: Description of service selection algorithm. Finally, the resulting data volume for the service mix is calculated as 4.3 Service Selection The last step in the description of the traffic generated by a user is the association of different services to the user. A flowchart illustrating the main process is presented in Figure 3. Initially, a total data volume (D Um ) and active probability (P Um ) during the busy hour is assigned, taking as reference the input distributions obtained from measurements. In this document, however, exemplary distributions are used for data protection purposes as explained in section 3. Afterwards, a combinatorial analysis is performed in order to select the possible services. For such a service mix with a total of N services, the busy probability per service is calculated based on [4]: P Um = j P (S j ) j<k P (S j S k )+ j<k<l P (S j S k S l ) +... + ( 1) N+1 j P (S j ), (4) where P (S j ) is the probability for service j of being active; P (S j S k ) is the probability of both services, S j and S k, being active. Services are assumed to be independent, therefore P (S j S k ) = P (S j )P (S k ). D Uc = N P (S j )R(S j ), (5) j where R(Sj) is the data rate for service j, defined from common codecs such as H.264 among others [6]. The resulting value is compared with the total user data volume assigned from input distributions and if the difference is minimal, the service mix is selected. In the case that there are several alternatives with total data volumes close to the desired value, the one with the lowest difference is selected. 5 Access Modeling To finish the user generation, a description of the access link qualities is given to the user. For the DSL access, line speeds from 1Mbps up to 16Mbps can be selected. For the LTE connection, an SINR distribution is applied, obtained from simulations in a System Level Simulator based on the 3GPP recommendation TR 36.942 [3], where the Shannon Bound is used to obtain a lookup table of data throughput per Physical Resource Block (PRB) vs. SINR. During the simulation, the number of scheduled resources in the LTE network varies, hence changing the total data rate a user can obtain from the wireless access.
5.1 Hybrid Access Algorithm After the users have been completely described, i.e. a service mix has been associated and the access specifications have been given, the simulation takes place in a second base manner. At every second (snapshot) the demand associated to the users is calculated from the current service states and the Hybrid Access Algorithm is executed to specify the way the traffic should be transmitted through the different accesses. The Hybrid Access Algorithm can be summarized as follows: 1. Identify the total demand per service type. 2. Assign the demand to the cheapest line first, i.e. DSL, until the maximum capacity is reached. 3. If there is remaining data to be transmitted, activate the second access (LTE) for resources scheduling. In the second step, different approaches can be used in order to assign the demanded data to the DSL link. For instance, a prioritization of the different service types can be performed to provide a better experience to the user. In the simulation, the LRT services have been prioritized to be only transmitted through DSL, thus the possibility of satisfying the user expectations is high. For the NLRT and NRT services, the remaining resources are distributed fairly. 5.2 Extended Round Robin Scheduler At the base station (BS) side, there are several algorithms to schedule the existing resources to the users. The TQoES uses an extended version of a Round Robin scheduler which takes also the user demand into account. In the first step, the BS schedules the radio resources in a round robin fashion, i.e. equal number of resources for each active user. Then, each demanding user can unlock the unused resources. The scheduler assigns the remaining resources also in a round robin fashion to users demanding more resources. The scheduling process ends either if there are no more resources to be scheduled, or each user has enough resources to serve their demand. To do not harm the existing mobile user, a prioritization in the scheduling process has been implemented. The mobile users have the highest priority and the BS assigns only the remaining resources to the Hybrid Access users. 6 Model Validation In the following, the correctness of the user modeling is verified. In Table 2 the simulation parameters used for the evaluation are given. Parameter Value No. radio cells 57 (3 cells per BS) Base station scheduler Extended Round Robin LTE Bandwidth 10 MHz BS Inter-site distance 5000 m Simulated time / granularity 1 hour / 1 s Environment Rural No. HA-user per radio cell 30 SINR distribution 3GPP 36.942 DSL capacity 1 Mbps, 3 Mbps, 6 Mbps Table 2: Simulation parameters 6.1 User Traffic Generation The initial step is to ensure that the user traffic generation was done correctly. Therefore, input and output distributions of the user demand and busy probability were compared. 1710 users (30 users per base station) with unlimited line speed to avoid any bandwidth limitation effects were generated. CDF CDF 1 0.8 0.6 0.4 0.2 0 1 0.8 0.6 0.4 0.2 input distribution output distribution 10 7 10 8 10 9 10 10 Demand [MB] input distribution output distribution 0 10 3 10 2 10 1 10 0 Busy probability Figure 4: Input and output distributions of the user demand (top) and busy probability (bottom) In Figure 4, the input and output distributions of the user demand and busy probability are depicted. The simulated distributions are very similar to the references. The limitation of the lowest busy probability (0.1 %) is the reason of the small deviation between the input and output distributions. Despite the small differences, the results validate the simulation to generate user traffic based on given distributions and due to that, it is used for further
evaluations. 1 7 Exemplary Results of Quality of Experience and Resource Demand The aim of the simulator is to evaluate the QoE enhancements introduced by the Hybrid Access Architecture under realistic user traffic conditions. In the following results are focused on NLRT and NRT services. In the implementation, LRT services are not influenced by the Hybrid Access Architecture due to their priority on the DSL and their low data rates. Comparison between Fixed and Hybrid access is presented, where Fixed access represents simulations with only DSL connections and, Hybrid access illustrates the simulations with both connections, DSL and LTE, active. 7.1 Non Linear Real Time Services The Quality of Experience for NLRT services is described by the ratio between good transmissions, i.e. successful streaming with average data rate over the defined threshold, and the total number of streamings. In Table 3 the percentages of good QoE for different DSL capacities are presented. By increasing the line capacity, the percentage of good QoE increases to 99.3 % with a 6 Mbps DSL. The Hybrid access significantly improves the QoE of the fixed line users with DSL capacities of 1 Mbps and 3 Mbps by 16.5 % and 18.7 %, respectively. DSL capacity [Mbps] 1 3 6 Good QoE Fixed [%] 79.2 80.8 99.3 Good QoE Hybrid [%] 95.7 99.5 100 Table 3: Percentage of good QoE for NLRT 7.2 Non Real Time Services NRT service QoE is described by the download rate. In Figure 5 the distributions of the NRT data rates are shown for DSL capacities of 1, 3 and 6 Mbps. 5 MB files are selected for the evaluation because they are transmitted very often and therefore, provide a reliable statistic. Fixed access users receive mostly with data rates corresponding to the full capcaity of the DSL. In some cases (< 10 %), the data rates are slightly lower due to the presence of other concurrently active services. Activating the Hybrid Access enhances the data rates, in average, by a factor of 2.3 to 7.5 for a 6 Mbps to 1 Mbps DSL access, respectively. That improvement would allow us to serve customers who suffer from low fixed line speeds with at least 7 Mbps in average. CDF 0.8 0.6 x7.5 x3.4 x2.3 FA-Users;1Mbps LS 0.4 HA-Users;1Mbps LS FA-Users;3Mbps LS 0.2 HA-Users;3Mbps LS FA-Users;6Mbps LS HA-Users;6Mbps LS 0 0 10 20 30 40 Data rate [Mbps] Figure 5: NRT data rate distributions for 5MB file size 7.3 Wireless Network Resource Demand The previously mentioned enhancements impact significantly the available resources of the wireless network, which also can be evaluated by the simulator. Therefore, we quantified the part of the data which is transmitted through the wireless network and the radio resource demand (see Tab. 4). We found that with increasing line speed the amount of data which is transmitted through the wireless network is decreasing. Hybrid access users with line speed of 1 Mbps transmitting about 66.5 % of their data through the wireless network, whereas users with 6 Mbps just transmitting 29.8 %. The simulator also enables us to quantify the transmitted data volume per radio cell and the utilized radio resources. We found that the data volume per radio cell decreases from 1.2 GB at 1 Mbps line speed to 0.5 GB at 6 Mbps line speed. Supporting the data demand of the hybrid access users with 1 Mbps line speed, the radio cells have to schedule in average about 36.6 % of their radio resources. Hybrid access users with 6 Mbps line speed need only 13.4 % of the radio resources. DSL capacity [Mbps] 1 3 6 Traffic share LTE [%] 66.5 44.4 29.8 Wireless data volume [GB/h] 1.2 0.8 0.5 Utilized radio resources [%] 36.6 23 13.4 Table 4: Data volume and utilized radio resouces per cell 8 Conclusions In this paper, we present a simulator to generate user traffic based on given user demand and busy probability distributions and to analyze the Quality of Experience of the users.
The SINR distribution mapped into a throughput distribution enabled us to quantify the possible data rates of the wireless network. Both, the user generated traffic and the wireless throughput distributions can then be used to evaluate the impact of the hybrid access architecture in a realistic manner. References [1] Cisco Visual Networking Index: Forecast and Methodology, 2012 2017, available at www.cisco.com [2] Deutsche Telekom AG: Capital Market Day, 2012, available at www.telekom.com [3] 3GPP: TR 36.942 v11.0.0, Radio Frequency (RF) system scenarios, 2012, available at www.3gpp.org [4] Sheldon M. Ross: Introduction to Probability Models, Tenth Edition, Elsevier, 2010. [5] Eugene Seneta: Non negative Matrices and Markov Chains, Second Edition, Springer Series in Statistics, 2006. [6] Kari Järvinen et. al.: Media coding for the next generation mobile system LTE, Elsevier Computer Communications, Vol. 33, pp.1916-1927, 2010.