Performance issues for VoIP call routing in a hybrid ad hoc office environment

Transcription

1 Performance issues for VoIP call routing in a hybrid ad hoc office environment Marc Portoles-Comeras, Josep Mangues-Bafalluy, and Marc Cardenete-Suriol Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) Parc Mediterrani de la Tecnologia (PMT) Av. Canal Olímpic S/N Castelldefels - Barcelona - Spain {marc.portoles,josep.mangues,marc.cardenete}@cttc.es Abstract This paper extends previous work on the impact of the number of hops on the quality of VoIP applications running over wireless multihop networks. The paper also proposes a framework for analyzing the performance of route re-discovery procedures in terms of the time that the communication is disrupted due to this process and the recency, or the time that it takes to the user to forget about the disconnection. The paper provides numerical and experimental data showing the relation between the voice codec used and the different design options that should be taken during deployment. Additionally, the paper proposes an application layer aggregation strategy that improves the maximum number of VoIP calls that can be supported in a wireless multihop network. Keywords-Wireless multihop networks, VoIP, voice codecs, data aggregation, extended E-model. I. INTRODUCTION Recent years have witnessed an extensive proliferation of wireless networking deployments. The low cost, high available bandwidth and the ever-increasing pervasiveness of communications strongly motivate the appearance of new applications and usage models for such networks. One of the key technologies taking advantage of this situation is voice over IP (VoIP). The growing popularity of companies and projects such as Skype is seamlessly introducing VoIP at all levels of the society as companies and end users are experiencing the cost advantages of integrating voice with their data networks. There have been repeated studies on the support of VoIP calls over IEEE based networks. However, despite the large available bandwidth, it can bee seen, experimentally, that only a limited number of calls can be supported with an acceptable voice quality [4]. This is even hardened in wireless multihop environments, where the mutual interference between the stations collaborating in forwarding a voice call reduces even more the number of calls supported (see [5]). Solutions have been proposed to control the number of calls admitted into an IEEE network [7], and in order to increase the total number of calls supported both in infrastructure networks [6] and wireless mesh environments [5]. An important issue when considering the utilization of VoIP in wireless environments is the routing protocol to be used. The way that the protocol deals with mobility is the key issue determining its performance in wireless infrastructure deployments (see [10] or [11]). However, when moving to multihop environments (e.g. ad hoc or mesh networks), the performance of a routing protocol directly depends on the scenario that it is to be applied to (see [3]). Designing a routing protocol able to support VoIP calls over a multihop environment is not straightforward. VoIP calls are very sensitive to packet losses and should be bounded on transmission delay to assure correct communication. On one side, packet losses in multihop environments are highly related to the saturation of the system [12]. The authors in [5], show how even a single VoIP flow traversing 6 wireless hops (transmitting, each, at 2Mbps physical rate) can completely occupy the bandwidth in a wireless network. The authors show also how a network layer aggregation strategy and a header compression strategy can fairly improve the resource utilization of VoIP calls, increasing then the number of calls supported in a wireless multihop network. In order to avoid undesirable throughput limitations, modern implementations of wireless ad hoc routing protocols [9] limit the maximum number of hops that a flow can traverse. They rely then on fixed networking infrastructure in a hybrid ad hoc network fashion to complete communication between all stations in the network [8]. However, determining the maximum number of hops to be traversed highly depends on the scenario, application and number of concurrent flows to be transported. On the other side, the end to end delay of a VoIP flow traversing a wireless multihop network is rarely an issue when flows traverse a reduced number of hops or when low-end devices are used. In this latter case forwarding nodes generally have short buffering capacity that rapidly overflows in case of saturation of the network, resulting in packet loss, not delay. Note also that jitter fluctuations are generally overcome using playout buffers at receivers. Finally, one should also consider the impact of disconnections on the quality of VoIP calls. Wireless multihop network paths, even in their more static form (e.g. all-wireless mesh networks), are prone to suffer disconnections. These disconnections trigger the route re-discovery process in the routing protocol to find a new appropriate path to support the communication. This may lead to packet losses and/or delays that affect the quality of ongoing VoIP calls. When designing a routing protocol for wireless multihop networks, the number of route rediscovery processes should be kept low and fast in order to support VoIP calls.

2 The study presented in this paper focuses (1) on extending previous work on the impact of the number of hops on the quality of VoIP calls and (2) providing a framework to analyze the impact of route disconnections on the quality of a VoIP as perceived by the end user. Results presented aim at providing meaningful results to guide the design of efficient strategies and protocols to support VoIP communications over wireless multihop networks. The paper presents several new contributions. Firstly, it shows how controlling the mutual interference between the multiple nodes collaborating in the forwarding of a single VoIP flow, unbounds the number of hops that this flow can traverse (as long as end to end delay is kept acceptable). Second, it shows how codec adaptation can help increasing the number of clients supported in a multihop collaborative environment. Third, it proposes an aggregation strategy at the application layer and shows how it effectively increases the number of VoIP flows supported in a wireless multihop network. Finally, it proposes a framework to analyze the impact of route rediscovery disconnections using extensions to the E-model [2]. VoIP quality is analyzed throughout the study using the E- model [1]. This technique maps network parameters to a subjective metric accounting for the quality of the phone call. Besides, the impact of the route rediscovery process is analyzed using the extended E-model [2], as it better accounts for the impact of burst losses caused disconnections. When the E-model is applied, an R-factor value characterizes the perceived quality of the call. The minimum acceptable call quality is obtained when the R-factor has a value of 70 (equivalent to the PSTN call quality.) This is our target value all throughout the paper. The rest of the paper is organized as follows. Section II defines the target scenario considered. Section III studies the impact of the number of terminals on the voice quality of the VoIP calls transported. Additionally, it proposes an application layer strategy to improve the number of terminals supported. Section IV proposes a framework to evaluate the performance of route re-discovery procedures in terms of the voice quality perceived by end users. Finally, section V concludes the paper. II. SCENARIO STUDIED: CALL RELAYING As mentioned before, studies in wireless multihop networks highly depend on the scenario considered. This section proposes a target scenario for the study. The results obtained may be extended to several other scenarios but generalization is beyond the purpose of this paper. Figure 1 presents the targeted scenario. A number of VoIP terminals collaborate in order to route individual VoIP calls between each one of the rest of terminals and peers residing in an external network. A wireless gateway is the one interconnecting the rest of the network and the terminals holding VoIP calls. There are a number of considerations to be taken into account to completely define the scenario. Firstly, all VoIP communications are established between a wireless terminal and a node that resides somewhere in the backbone side of the network in the figure. Secondly, all terminals support establishing a single VoIP communication but are able to forward VoIP calls from other terminals. This can be thought as the communications infrastructure inside an office, where, in order to reduce costs, a single wireless gateway is installed and its coverage is extended through collaborative relaying. Third, the mobility of terminals is reduced. In general, users holding each one of the terminals remain in the same position for a reasonably long time. However, they might occasionally move, which leads to triggering a route re-discovery process. Fourth, VoIP terminals use the IEEE WLAN protocol to form the multi-hop network. Finally, the mean duration of a call is considered to be of 2.6 minutes, as reported in [13]. Figure 1. Scenario considered The following section presents a study of the impact of the number of terminals and hops on the quality of the VoIP calls, while section IV studies the impact of route re-discovery latency time on the quality of the VoIP traffic. The first study is eminently experimental while the other presents a theoretical analysis framework. III. IMPACT OF THE NUMBER OF TERMINALS ON THE VOICE QUALITY This section analyzes the relation of the number of terminals involved in the communication and the number and quality of VoIP calls that can be supported. It also proposes an application layer aggregation strategy to increase the number of VoIP calls supported in a network. A. Experimental setup All experiments have been carried out within EXTREME framework (see [16]). This is a multi-purpose networking experimental platform currently under development within the Centre Tecnològic de Telecomunicacions the Catalunya (CTTC) in Barcelona. The main advantage of this platform is its high automation capabilities that allow automatic execution, data collection and data processing of several repetitions of an experiment. All computers involved in this scenario are Pentium IV PCs with 512MB of RAM memory. They all run Linux operating system with Kernel Figure 2 shows the experimental setup used to obtain the results described in this section. Several nodes of the EXTREME cluster are each equipped with a PCI based wireless card carrying the popular Prism chipset (specifically, Z-COM ZDC XI-626 WLAN cards). All wireless cards are interconnected using coaxial wiring and propagation losses are emulated using RF attenuators, as depicted in the figure. This serves a double purpose. Firstly, it isolates the

3 experimentation environment from external interferences. Secondly, it allows us to build a highly controlled environment where transmission range and carrier sense range can be adjusted in accordance to our needs for each one of the stations. Specifically, we adjust attenuations so that each one of the stations can only transmit to its neighboring ones but can sense those stations that are two hops away. When required wireless terminals start UDP traffic flows emulating VoIP traffic and with destination the VoIP sink in the figure. At the same time the VoIP sink starts an equal flow in the reverse direction completing the bidirectional VoIP communication. VoIP flows are emulated using the MGEN tool [14]. The reason to choose this application is double fold. On one side the traffic source can activate an option called precise on that efficiently controls real-time generation of packets. On the other side the traffic sink is able to store received packets in a format that allows convenient packet loss count and latency computation afterwards. Figure 2. Experimental setup B. Impact of the number of hops on a single-flow VoIP quality Here we analyze the impact of the number of hops traversed on the quality of a VoIP call. Figure 3 shows the R-factor obtained when computing the E-model at the VoIP sink when the terminal it communicates with is located from 1 to 7 hops away. Note that, while it is a bidirectional communication, we do not print the curve obtained at the wireless terminal, as it is practically the same as the one obtained at the VoIP sink node. Note also, that all nodes are configured to transmit at 2Mbps physical rate. terminals (those within carrier sense range), so its available channel resources (bandwidth) will, at most, be divided by four [15]. The resulting per node throughput capacity is constant and sufficient to support a VoIP call, no matter the number of hops packets have to traverse. As a result, one can say that a conscious interference-aware deployment of a wireless multihop network can help, in the absence of other background traffic, increasing the maximum number of hops that a VoIP flow can traverse without suffering any quality degradation. C. Impact of the number of hops on multi-flow VoIP quality In this case we study the quality of voice conversations when each one of the terminals present in the network starts a VoIP call with the VoIP sink node. Going back to Figure 2, each one of the terminals communicates with the VoIP sink node via the neighboring node closest to the wireless gateway, which relays all VoIP messages in both directions. The idea is to study the maximum number of terminals supported in such a scenario. Note that this is a chain topology, so that there is only one route possible from each one of the terminals and the VoIP sink node. Figure 4a and Figure 4b, plot the quality of the VoIP call between the last terminal and the VoIP sink, both at the VoIP sink node and the wireless terminal respectively. They plot the voice quality versus the total number of terminals (clients) maintaining a VoIP conversation with the VoIP sink node. Note that the number of terminals is equivalent to the number of hops traversed by the VoIP flows going to (and from) the last wireless terminal. Figures show how, when using the G.711 codec, the system can only sustain up to 5 VoIP calls in the scenario described with an acceptable quality (R>70). This raises when using G.729 or GSM codecs, as the system can sustain up to 6 calls. Figure 3. R-factor vs number of hops traversed by a single flow as observed at the VoIP sink node This figure differs from the results reported in [5] as here a single VoIP flow can be sent over a larger number of hops. The main reason for this lies in the following observation. With the interference model adopted here (carrier sense range only reaches two hop distance) we assure that, at any time, any node only contends for channel access with at most four other (a) (b) Figure 4. R-factor of the last terminal call, as observed (a) at the VoIP sink and (b) at the wireless terminal for a different amount of active terminals Two observations can be raised from the figures. Firstly, the breakdown previously reported in [12] is shown experimentally. When the network saturates, the R-factor suffers a sudden breakdown preventing any communication between the last node and the VoIP sink. Secondly, when using different voice codecs, a different number of hops is reached. This suggests that a voice codec adaptation might be an efficient strategy to support a higher number of voice calls in saturated wireless multihop environments. It is worth mentioning here, that not only communications with the last terminal in the chain but practically all the rest of

4 the VoIP communications fall into unacceptable voice quality situation when the network enters saturation (i.e. when the last terminal starts VoIP transmission). This change is also abrupt, in a breakdown manner, which challenges the design of admission control mechanisms. D. Proposed application layer aggregation of packets and its impact on the number of VoIP calls supported In order to increase the number of VoIP calls supported in a mesh networking deployment, the authors in [5] propose using an aggregation strategy to be applied at the networking stack of each one of the wireless mesh nodes. While this strategy presents promising results, some considerations should be taken, regarding its application in our targeted scenario. Firstly, this solutions implies modifying the networking stack of all the terminals to be used in order to include the proposed algorithm. This modification must be done at the OS level which increases complexity of the task. Secondly, considering that short buffering capacity is expected in the relaying terminals, no much forwarding opportunities might arise for packet aggregation. Here we propose, as an alternative, doing the aggregation at the VoIP application itself. When the quality of the call being maintained is detected to have poor quality (e.g. through RTCP notification and run-time R-factor computation) the application can alternatively choose to aggregate various voice packets into one, prior to the send process. This process reduces the amount of resources required to keep the communication. The number of packets to be sent is reduced and this leads to reducing the amount of overhead to send them. This strategy has, however, an impact on the end-to-end delay of packets. In order to conduct aggregation some packets are delayed in purpose. However, as explained above, the end-to-end delay is not, generally, an issue in the targeted scenario, so there exists a margin of tolerance. (a) (b) Figure 5. R-factor of the last terminal call, as observed (a) at the VoIP sink and (b) at the wireless terminal for a different amount of active terminals Figure 5a and Figure 5b show similar plots as those in Figure 4. In this case, however, stations are applying the aggregation strategy proposed. Each one of the terminals aggregates at the application layer two VoIP packets into one and sends them together to the next hop towards the VoIP sink node. The extra delay suffered by some packets due to the aggregation process is accounted for in the computation of the R-factor value. However one might notice that as the end-toend delay is still low (<150ms) the R-factor value does not reflect any change. The figures show, however, how the aggregation effectively serves the purpose of supporting a higher number of active VoIP terminals in the network chain. This results suggest the possibility of including aggregation strategies at the application layer instead of the lower layers, as this extends the maximum number of terminals supported in our target scenario. IV. IMPACT OF ROUTE RE-DISCOVERY LATENCY TIME ON THE VOICE QUALITY This section proposes a framework to analyze the impact of the route re-discovery process on the quality of the voice call as perceived by the end-user. The framework is based on using the extended E-model [2]. Next subsections detail why the extended E-model is more appropriate to analyze the impact of this process and propose a framework to determine the performance of a devised re-discovery process taking into account the disconnection time and the length of the VoIP call. A. The extended E-model is more appropriate for transient disconnections The E-model, used to compute the R-factor throughout the paper, assumes that packet losses are uniformly distributed over time. It is then, not appropriate to account for the impact of route rediscovery disconnections that lead to bursty losses and transient delays. The extended version of the E-model [2] incorporates some time-varying impairments that are not considered within the E-model. One of them refers to the fact that the subjective quality perceived by users changes more slowly than the quality calculated by using the instantaneous packet loss and other impairments. The transitions between burst (period of time during which a high percentage of packets are lost) and gap (the packet loss rate is very low) states are corrected by using exponential decays with time constants of 5 seconds for the gap to burst transition and 15 seconds for the burst to gap transition [2]. Another effect is the recency, which is based on the fact that people tend to remember the most recent events. It is modelled using an exponential decay in the perceived quality which starts at the end of the last significant burst of packet loss and approaches the average quality level for the call. B. A framework to analyze route re-discovery disconnections The bursty loss resulting from the transient disconnection suffered by the VoIP terminal during the route re-discovery process and the recency of the user after reestablishing regular communications are factors to be considered to analyze the appropriateness of a route re-discovery process. Figure 6 plots the R-factor value perceived by a VoIP versus the time elapsed since the route discovery disconnection finished. This is plotted for various disconnection times, ranging from 200ms (typical in infrastructure based WLAN networks) and 5 seconds (a value considered well beyond acceptance for real time communications). Plotted curves show that when the disconnection time is below one second the user does not perceive unacceptable quality degradation. Even when the disconnection takes around 2 seconds the user forgets about the disturbance at about 15 seconds after the VoIP communication is reestablished. Note, however that this

5 values do not account for mean end-to-end packet losses and delays that should be included for completeness in the curve. Observing the curve one can notice that a long disconnection is preferable to several shorter frequent ones, as the user may rapidly forget about a single disconnection but would not tolerate frequent shorter ones. Once a protocol and route rediscovery have been designed, curves in this plot may serve to evaluate the possibility to support quality VoIP calls. Figure 6. Impact of route rediscovery disconnections on the quality of VoIP calls when using the G.711 codec Figure 7. Impact of route rediscovery disconnections on the quality of VoIP calls when using the G.729 codec For completeness, Figure 7 plots similar curves for the G.729 codec case. When using this codec the user is less tolerant to disconnection times and a maximum of 1 second occasional disconnections is tolerated after which it takes around 15 seconds for the user to forget about the annoyance. These plots suggest again the use of codec adaptation strategies in order to adapt the communication to the network conditions. While G.729 might be more attractive in order to support a higher number of calls in a network, it is less recommended when the route re-discovery process incurs high latency. V. CONCLUSIONS The study presented in this paper extends previous work on the impact of the number of hops on the quality of VoIP calls and provides a framework to analyze the impact of route disconnections on the quality of a VoIP as perceived by the end user. Results presented aim at providing meaningful results to guide the design of efficient strategies and protocols to support VoIP communications over wireless multihop networks. The paper presents several new contributions. Firstly, it shows how controlling the mutual interference between the multiple nodes collaborating in the forwarding of a single VoIP flow, largely extends the number of hops that this flow can traverse, without a significant penalty on the perceived voice quality. Second, it shows how codec adaptation can help increasing the number of clients supported in a multihop collaborative environment. Third, it proposes an aggregation strategy at the application layer and shows how it effectively increases the number of VoIP flows supported in a wireless multihop network. Finally, it proposes a framework to analyze the impact of route rediscovery disconnections using extensions to the E-model [2]. Through this framework the study shows also how code adaptation might be also useful to enhance the user experience in case of route re-discovery disconnections. ACKNOWLEDGMENTS This work was supported in part by the European Commission project WIP under contract and by Generalitat de Catalunya under grant number SGR (Grup de Recerca Singular). REFERENCES [1] ITU-T. "The E-model, a computational model for use in transmission planning" ITU-T Recommendation G.107, May 2000 [2] Clark, Extensions to the E-model to incorporate the effects of time varying packet loss and recency, TIA standard T1A1.1/ , 2001 [3] T. Camp, J. Boleng, and V. Davies, A Survey of Mobility Models for Ad Hoc Network Research, Wireless Comm. and Mobile Computing (WCMC), vol. 2, no. 5, pp , [4] Anjum, F.; Elaoud, M et al.., "Voice performance in WLAN networks - an experimental study," IEEE GLOBECOM '03, Dec [5] S. Ganguly et al., Performance Optimizations for Deploying VoIP Services in Mesh Networks, in IEEE JSAC, November [6] P. Clifford et al., On improving voice capacity in infrastructure networks, in IEEE WirelessCom 2005, Maui, Hawaii, USA, [7] Garg, S.; Kappes, M., "Can I add a VoIP call?," in proc. of the ICC '03. May 2003 [8] B. Xu, S. Hischke, and B. Walke, The Role of Ad Hoc Networking in Future Wireless Communications, Proceedings of ICCT [9] C. Tschudin, R. Gold, O. Rensfelt and O. Wibling: "LUNAR: a Lightweight Underlay Network Ad-hoc Routing Protocol and Implementation." Proc Next Generation Teletraffic and Wired/Wireless Advanced Networking (NEW2AN'04), St. Petersburg, Feb 2004 [10] M. Portoles-Comeras, M. Cardenete-Suriol, J. Mangues-Bafalluy, M. Requena-Esteso, L. Hersant, "Experimental assessment of VoIP quality in MIPv6 and SIP mobility scenarios", in proceedings of the ICC 2006, Istanbul, Turkey, June, 2006 [11] M. Cardenete-Suriol, J. Mangues-Bafalluy, M. Portoles-Comeras, M. Requena-Esteso, M. Gorricho, "VoIP performance in SIP-based vertical handovers between WLAN and GPRS/UMTS networks", to appear in proc. of ICC 2007, Glasgow [12] P. Falconio et al., "Performance of a multi-interface based wireless mesh backbone to support VoIP service delivery", in proceedings of the WiMob 2006, Montreal, Canada, June, 2006 [13] D. Lam, D. D. Cox, J. Widom, Teletraffic Modeling for Personal Communications Services, IEEE communications magazine [14] The Multi-Generator Toolset (MGEN) available at [15] J. Jangeun; M.L. Sichitiu, "The nominal capacity of wireless mesh networks," IEEE Wireless Communications, vol.10, no.5, Oct 2003 [16] M. Portoles-Comeras et al., "EXTREME: Combining the ease of management of multi-user experimental facilities and the flexibility of proof of concept testbeds", in proc. of the IEEE TridentCom 2006.