Performance Evaluation of Access Selection Algorithms for VoIP on Wireless Multi-Access Networks

Transcription

1 VI INTERNATIONAL TELECOMMUNICATIONS SYMPOSIUM (ITS26), SEPTEMBER 3-6, 26, FORTALEZA-CE, BRAZIL Performance Evaluation of Access Selection Algorithms for VoIP on Wireless Multi-Access Networks A. P. da Silva, F. de S. Chaves, V. A. de Sousa Jr., R. A. de O. Neto and F. R. P. Cavalcanti Abstract The tendency of wireless communication systems beyond 3G is the wireless Multi-Access (MA) environment, which is characterized by the integration of different Radio Access Networks (RANs). Parallel to the wireless network evolution, Voice over Internet Protocol (VoIP) is one of the fastest growing Internet real-time applications. Implementation of VoIP over a Multi-Access network requires the incorporation of efficient resource allocation strategies, namely, Common Radio Resource Management (CRRM), in order to meet the optimum combined network performance. In this paper, we propose Access Selection (AS) algorithms for MA networks and evaluate their performance for VoIP service. Conventional AS algorithms are taken as reference. Index Terms Access Selection, Multi-Access Networks, VoIP, systems beyond 3G. I. INTRODUCTION In Multi-Access (MA) networks, the availability of different Radio Access Networks (RANs) makes possible the selection and the combination of several application services. These scenarios are characterized by the composition of wireless networks that is the focus of future wireless communication systems beyond 3G. This kind of wireless network becomes a very attractive paradigm in the sense of high-quality, since the mobile services differences with respect to cost and resource utilization are significative. To provide multimedia services over the upcoming wireless MA networks, intelligent and advanced Radio Resource Management (RRM) techniques may be extend to their context. This new facet of resource management requires a new coordination entity in order to manage the radio resources available in each RAN. Then, the CRRM, now acting as a system entity, plays the role of a policy manager for the access to the radio resources in the MA networks. It is responsible for tasks like the initial user access selection, the user mobility among RANs and the network level congestion control, among others. An important RRM function of the CRRM is the AS procedure. In the CRRM context, the AS procedure has as an objective the selection of the best RAN in the composite network for the admission of a new connection. In this case, user s characteristics as type of service and cost preferences as well as the status of the systems which compose the network may be considered in the decision process. The authors are with the GTEL-UFC: Wireless Telecom Research Group, Teleinformatics Engineering Department, Federal University of Ceara, Brazil. URL: {alex, fabiano, vicente, neto, rodrigo}@gtel.ufc.br. Following the internet evolution, VoIP is one of the fastest growing services. It is an interesting alternative to the Public Switched Telephone Network (PSTN) voice services because it brings some economic and technical advantages. A great technological challenge will be to convey voice over wireless MA networks retaining the Quality of Service (QoS) that users associate with traditional telephone networks. In the specialized literature, both MA network and VoIP services are not approached in a combined manner. In [1], VoIP capacity is studied for High Speed Downlink Packet Access (HSDPA). In [2], a VoIP admission control strategy based on a channel utilization estimation is proposed for the IEEE 82.11b wireless network. In both works, VoIP studies are not carried out under wireless Multi-Access scope. In [3], Yilmaz and Furuskär propose and evaluate a set of access selection algorithms in a multi-access scenario, but the investigations are exclusively for best-effort data services and the dynamism of the wireless systems is not stressed. This paper investigates the performance of access selection algorithms for VoIP in a dynamic multi-access scenario. In order to make the following discussions clearer and the computational simulations simpler, we consider that the Multi-Access system is composed by two Radio Access Networks. Albeit there is no generality losses with respect to the ideas, the developed algorithms can be applied to networks composed by a higher number of individual RANs. The rest of the paper is organized as follows. In section II, we present the VoIP and Multi-Access network modeling used in this paper as well as their relationships. Discussions about the Access Selection strategies and their definitions are carried in section III. Section IV demonstrates the performance of the proposed AS algorithms by means of simulation campaigns. Finally, the conclusions and perspectives are summarized in section V. II. MODELING OF VOIP ON MULTI-ACCESS NETWORKS The convergence of packet-switched systems and voice service introduced a new challenge with the fact that packet-switched system were not conceived to deal with QoS-guaranteed services. This feature should be inserted in the system model, mainly those aspects directly related to delay throughout the transmission chain. This section presents the most important aspects that characterize the VoIP service and the RANs that will be used to assess our Access Selection proposals. SBrT 433

2 VI INTERNATIONAL TELECOMMUNICATIONS SYMPOSIUM (ITS26), SEPTEMBER 3-6, 26, FORTALEZA-CE, BRAZIL A. VoIP Modelling 1) User Arrival Model: Voice calls arrive according to a Poisson process with a specific arrival rate. The voice load is determined by the average time between consecutive calls and the average call duration, both exponentially distributed. 2) Voice traffic model: A speech source creates a pattern of talkspurts and gaps, as classified by the speech activity detector [4]. There are principal spurts and gaps related to the talking, pausing, and listening patterns of a conversation. There are also minispurts and minigaps due to short silent intervals that punctuate continuous speech [4]. We assume a slow activity detector that responds only to the principal talkspurts and gaps. Then, our voice traffic model following a traditional approach in which the call duration is assumed to be an exponential distributed random variable with mean of 9 seconds. The VoIP call is composed by activity and silence periods with equal probability. The duration of both activity and silent periods are exponentially distributed with a mean of 3 seconds [5]. During activity periods, we assume a Constante Source Rate model in which a voice frame with 576 bits is generated every 2 ms, corresponding to a rate of 28.8 kbps. Figure 1 illustrates our model. The frame includes the overhead bits by the addition of /UDP/IP headers (16 bits of IPv bits of UDP + 96 bits of ), tail (12 bits) and payload (244 bits) [6]. Activity Period IPv4 16 bits UDP 64 bits Silent Period 96 bits Payload 244 bits Tail 12 bits Fig. 1: Illustration of VoIP traffic Modeling. 3) Link Layer model: According to [1], in order to achieve an acceptable quality for the VoIP call, the one-way mouth-to-ear delay should be less than 25-3 ms. This total delay should account for all the nodes in the communication path: the voice codec, local access network or PSTN, Internet, wireless core network, base station and mobile equipment. The present paper modeled a delay budget inside individual RANs by means of a packet (frame) discard mechanism at the Medium Access Control (MAC) level. Studies conducted in [1], [8] [11] considered delay budgets in the range from 8 ms to 15 ms. This range should be sufficient for scenarios where the VoIP call is between two mobiles or between a land-line and a mobile user. Therefore, a maximum delay of 1 ms is considered in the packet discard model, i.e., generated frames with waiting time higher than 1 ms are assumed lost. B. Multi-Access Model 1) Deployment Model: For the performance result evaluation, we model a multi-access scenario composed by two generic Radio Access Networks (RANs), namely, RAN 1 and RAN 2. Each RAN emulates the data transmission over shared channel and all VoIP users are able to connect to both RANs. A user transmitting in a given RAN does not produce interference to the other RAN, because we assume that the RANs use different frequency bands. Neither user blocking nor vertical handover are assumed in order to ensure that users are always assigned to one of the RANs and they stay connected to one RAN up to the end of their sessions. We evaluate the performance of AS algorithms in a homogeneous deployment environment, i.e., the coverage area is the same for both RANs and the users are uniformly distributed over the coverage area as a whole. 2) Wireless system model: In the wireless system model, the effect of the channel is perceived through the radio link quality, expressed by the Signal-to-Noise Ratio (SNR). The SNR is modeled as a Gaussian distributed random variable with mean SNR mean and standard deviation σ. The user transmission rate is given by a link adaptation curve. We used an idealized linear mapping curve, which has for the input parameters the minimum and maximum values of rates and their respective SNRs. More details about this link adaptation model can be found in [12]. All above mentioned parameters make up a generic radio access network modeling which is used to characterize a homogeneous deployment MA scenario. Another aspect is assumed for the tuning of such parameters in each RAN: the user mobility. We emulate the mobility effects by means of a SNR variation model. Periodically, independent SNR samples are generated for all users obeying a path gain decorrelation distance and considering pedestrian mobility pattern with speed mean of 3km/h, i.e., users does not roam around the network, but their radio link qualities (SNR) vary as the simulation evolves in time. In traditional wireless networks, Real Time (RT) services (e.g., voice) are carried over dedicated channels because of their delay sensitivity, while the Non-Real Time (NRT) ones (e.g., web browsing) are transported over time-shared channels because of their burstiness. In accordance with VoIP premise, it has recently been proposed that even RT services can be efficiently transported over time-shared channels. In order to model this feature, the generated voice frames are put into a queue and delivered obeying some scheduling strategy. Here, the Round Robin scheduling policy is adopted for the queuing system because its inherent fairness quality. We also assume a time resolution of 2 ms to the voice packets assignments. This corresponds to the Transmission Time Interval (TTI), i.e., the time that the Round Robin (RR) algorithm can assign a different user to the transmission medium. C. Performance Metrics 1) VoIP delay: The delay of the VoIP service considered in this work regards only to the radio access network (transmission delay), i.e., due to scheduling and the users queuing. 2) Frame Erasure Rate: The Frame Erasure Rate (FER) is defined as the ratio of the number of lost frames to the total of frames. A VoIP frame loss can occur due to either the effect SBrT 434

3 VI INTERNATIONAL TELECOMMUNICATIONS SYMPOSIUM (ITS26), SEPTEMBER 3-6, 26, FORTALEZA-CE, BRAZIL of channel errors (poor SNR) or shared channel overload. A VoIP frame is discarded when it arrives at the scheduler too late, i.e., its delay is higher than a delay budget threshold. 3) User satisfaction: The user satisfaction is reached if the service provided fulfills its requirements. A VoIP user is assumed satisfied if it has a FER lower or equal to 2%, reflecting a good perceived voice quality provided by the speech codec [13]. III. ACCESS SELECTION ALGORITHMS Five Access Selection strategies were evaluated in this work. Two of them were originally investigated in [14], with the resulting Load Balancing Algorithm () and Rate Maximization Algorithm () taken as reference. We propose other three AS algorithms. A comprehensive description of all of them is listed in the following. A. Load Balancing Algorithm () The strategy is based on both the offered load and maximum transmission capacity in all RANs. According to the, the assignment of a new user must be made so that the normalized load of both RANs becomes closer: Used Capacity at RAN1 Used Capacity at RAN2 Total Capacity of RAN1 Total Capacity of RAN2 Then, a new user is connected to the RAN with lower normalized load. B. Rate Maximization Algorithm () In the rate maximization strategy, the user transmission rate is estimated in both RANs. The estimate is based on the user signal quality and on the link capacity of each RAN, i.e, the user rate is a mapping of current user SNR using the link adaptation curve (see section II-B). A new user is then admitted to that RAN where it experiences the highest transmission rate. C. Minimum Transmission Delay Algorithm () In the minimum transmission delay strategy the average user s transmission delay is considered as the decision metric. This metric is estimated taking into account only the users who had already transmitted at least a frame. When a specific user gets the channel and set of frames is transmitted, the transmission time (delay) of each frame is computed, making possible the estimation of the average transmission delay for this set of frames. In another user s transmission opportunity, all past transmitted frames and the current transmitted ones are taken into account to update the estimation of delay metric. We can note that the lost frames are not considered in delay computation. Then, the average transmission delay per user information allows the calculation of the average transmission delay for each RAN. The new user is admitted to the RAN where it experiences the lowest RAN delay. D. Minimum FER Algorithm () The minimum FER strategy is based on the user FER. The average FER can be determined for each RAN. The new user is admitted to that RAN where it experiences the lowest average FER. E. Utility-based Algorithm () The utility-based strategy takes into account information of both the new user and the network conditions. It is based on a utility function defined as the ratio of the mapped link bit rate of the incoming user to the user average transmission delay and calculated for each RAN. is a combined proposal of the and algorithms. The utility function provides a good satisfaction indicator when a low average transmission delay and high bit rate is perceived. The higher the mapped link bit rate of the incoming user in a given RAN and/or the lower the average transmission delay observed in that RAN, the higher the utility of the new user in the given RAN. Therefore, the new user is admitted to that RAN where it experiences the highest utility. IV. PERFONCE RESULTS This section presents the performance results of the AS algorithms. The VoIP performance is measured in terms of FER, average user transmission delay and user satisfaction (see section II-C). and are used as the reference algorithms. We analyze all algorithms under the scenario where the maximum capacities of the RANs are the same (R max1 = 2Mbps and R max2 = 2Mbps) and the user are uniformly deployed over the system as a whole, i.e., the homogenous deployment case. Heterogeneous deployment structure and different RAN capacities are not considered. Figure 2 shows the Cummulative Distribution Function (CDF) of the average transmission delay of the users, considering a request rate of 1 req/s. Apart from the algorithm, all other algorithms present practically the same performance. presents the best delay performance among all algorithms. In spite of the algorithm is based on the delay, at least two reasons can explain why this strategy does not present a better performance than the other ones. First, the update of the AS metrics used for the algorithms is performed at each 1 ms. This period is relatively long if compared to the frame transmission delay of each user. Second, the average transmission delay metric is only updated at the end of the activity periods, not taking into account the transmission delay variability at the frame level. The low performance of the is caused by the relatively long update period of the AS metrics and also due to the updating instant of the FER metric, which occurs at the end of the activity periods. These results indicate that the usage of delay and FER as an AS criterion provides a poor performance if we consider their average values. Then, the medium-term action performed by the proposed AS criteria must be based on time-restricted measurements. Besides, the described equal-capacity scenario is well-behaved. SBrT 435

4 VI INTERNATIONAL TELECOMMUNICATIONS SYMPOSIUM (ITS26), SEPTEMBER 3-6, 26, FORTALEZA-CE, BRAZIL Average Transmission Delay (R = 2 Mbps and R = 2 Mbps) max1 max Fig. 2: CDF of Average Transmission Delay (Request Rate equal to 1 req/s). Fig. 4: Average Transmission Delay: 9th percentile and mean..2.1 User FER (R max1 = 2 Mbps and R max2 = 2 Mbps).1.5 Fig. 5: User FER: 9th percentile and mean. Fig. 3: CDF of User FER (Request Rate equal to 1 req/s). Figure 3 shows the CDF of the user FER considering a request rate equal to 1 req/s. Similarly to the previous analysis, all algorithms present practically the same performance, except due to the same reasons explained before. We observe that the algorithm present the better performance if compared to the other ones. In order to bring complementary information of delay and FER, the figures 4 and 5 shows the 9th percentile and the mean of both the average transmission delay and the user FER for different offered loads (.2,.33,.5,.67, 1., 1.25, 1.33, 1.43, 1.67 and 2. req/s). The,, and algorithms present the same performance for average transmission delay in both 9th percentile and mean curves. In both user FER curves, the algorithm presents a subtly better performance for the range of request rates considered. Figure 6 shows the satisfaction of users for the same range of request rates. A user is considered satisfied if its FER is smaller than the maximum admissible FER, considered in our simulation to be 2 %. This performance metrics aim at the reflection of delay and FER behaviors in a unified metric. Similar to the previous results, the and have the same performance considering the homogeneous deployment MA scenario and the AS decision chosen as the average of proposed criteria Satisfaction of users (R max1 = 2 Mbps and R max2 = 2 Mbps).2 Fig. 6: Satisfaction of users. SBrT 436

5 VI INTERNATIONAL TELECOMMUNICATIONS SYMPOSIUM (ITS26), SEPTEMBER 3-6, 26, FORTALEZA-CE, BRAZIL V. CONCLUSIONS AND PERSPECTIVES This work has evaluated performance of AS algorithms in two generic equal-capacity RANs in a multi-access network under a homogeneous deployment scenario for VoIP services. Five AS algorithms were investigated:,,, and. Firstly, we assessed the CDF of the average delay and the CDF of the user FER for a single load: 1 req/s. For that load, we can indicate that and have presented the best performance in terms of average transmission delay and user FER respectively. We have also evaluated all algorithms for a different range of loads. Apart from the, we observed that all algorithms presented the same performance for both the 9th percentile and the mean for the average transmission delay. In terms of user FER, the good performance of the algorithm could be perceived for some range of loads. Finally, we conclude that the medium-term action performed by the proposed AS criteria must be based on more time-restricted measurements instead of the average. As perspective, we intend to investigate the relation of the update period of AS criterion metrics and the AS performance. This way, we expected to improve the performance of and algorithms. We also intend to study the joint operation of the AS and other CRRM techniques, for instance, the vertical handover. This is an interesting research topic in the sense of that the vertical handover can monitor the AS criterion in order to improve the overall system capacity. Finally, the conception and evaluation of CRRM algorithms regarding the coexistence of VoIP and best-effort services will be a great challenge in terms of resource management and QoS satisfaction. [7] B. Wang, K. I. Pedersen, T. E. Kolding, and P. E. Mogensen, Performance of VoIP on HSDPA, IEEE 61st Vehicular Technology Conference (VTC Spring), May 25. [8] M. Ericson, L. Voigt, and S. Wänstedt, Providing Reliable and Efficient VoIP over WCDMA, Ericsson Review White Paper, no. 2, pp , 25. [9] P. Hosein, Capacity of Packetized Voice Services over Time-Shared Wireless Packet Data Channels, IEEE INFOCOM, March 25. [1] W. Xiao, A. Ghosh, D. Schaeffer, and L. Downing, Voice over IP (VoIP) over Cellular: HRPD-A and HSDPA/HSUPA, IEEE 62nd Vehicular Technology Conference - VTC Fall, September 25. [11] E. J. Hernandez-Valencia and M. C. Chuah, Transport Delays for UMTS VoIP, IEEE Wireless Communications and Networking Conference - WCNC, no. 1, pp , September 2. [12] V. A. de Sousa Jr., R. A. de O. Neto, F. de S. Chaves, L. S. Cardoso, J. F. Pimentel, and F. R. P. Cavalcanti, Performance of access selection strategies in cooperative wireless networks using genetic algorithms, World Wireless Research Forum (WWRF), Paris, France, Tech. Rep., 25. [13] 3GPP, Performance Characterization of the Adaptive Multi-Rate (AMR) Speech Codec, 3 rd Generation Partnership Project, Sophia Antipolis, France, Tech. Rep. TR v6.. - Release 6, December 24. [Online]. Available: [14] O. Yilmaz, Access Selection in Multi-Access Cellular and WLAN Networks, Master s thesis, Royal Institute of Technology, Sweden, Stockholm, February 25. ACKNOWLEDGEMENT This work is supported by a grant from Ericsson of Brazil - Research Branch under ERBB/UFC.1 Technical Cooperation Contract. The authors A. P. da Silva, V. A. de Sousa Jr. and R. A. de O. Neto are scholarships supported by FUNCAP and Francisco R. P. Cavalcanti was partly funded by CNPq - Conselho Nacional de Desenvolvimento Científico e Tecnológico, grant no /22-8. REFERENCES [1] B. Wang, K. I. Pedersen, T. E. Kolding, and P. E. Mogensen, Performance of VoIP on HSDPA, IEEE Vehicular Technology Conference, June 25. [2] S. Garg and M. Kappes, Admission Control for VoIP Traffic in IEEE Networks, IEEE Global Telecommunications Conference, December 23. [3] O. Yilmaz, A. Furuskar, J. Pettersson, and A. Simonsson, Access Selection in WCDMA and WLAN Multi-Access Networks, IEEE Vehicular Technology Conference, 25. [4] D. Goodman and S. Wei, Efficiency of packet reservation multiple access, IEEE Transactions on Vehicular Technology, vol. 4, pp , [5] UMTS, Selection Procedures for the Choice of Radio Transmission Technologies of the UMTS, ETSI, UMTS TR v.3.2., Tech. Rep., April [6] 3GPP, Packet Switched Conversational Multimedia Applications: Transport Protocols, 3 rd Generation Partnership Project, Sophia Antipolis, France, Tech. Rep. TS v6.. - Release 6, June 24. [Online]. Available: SBrT 437