1 Performance evaluation of voice handovers in real 802.11 networks Bj¨ orn Gr¨ onvall and Ian Marsh CNA Lab. SICS Sweden {bg, ianm}@sics.se Abstract— In this paper we look at the performance of 802.11 networks for real-time voice communication. In particular, we focus on triggering timely handovers of the voice calls from the 802.11 network to the cellular infrastructure. We have devised an algorithm that uses link layer, IP and application measurements to determine when best to handover an ongoing call. Hence, determining the optimal handover time is the research challenge we face. A unique aspect of this challenge is that either poor algorithmic, or network performance, will be audible as poor voice quality for the end user. Our solution must also be robust and flexible due to the varying environmental conditions and differing hardware types. We look at two typical problem scenarios for 802.11 networks, one is poor coverage situations, and the other is sudden heavy traffic loads. We will show that it is possible to address the coverage problem, but heavy load situations are difficult to solve with a handover solution. We include our experiences whilst designing and evaluating this solution, plus give hints to the software and firmware writers of 802.11 products. I. I NTRODUCTION We have designed, implemented and evaluated a solution to allow the seamless roaming of voice calls between 802.11 and the cellular network. The focus was on the performance and ability of the 802.11 network to 1) provide sufficient network information to the application so that it can timely select access for the voice calls. 2) provide sufficient network resources for real-time voice calls. This work assumes there is another network technology available in the same physical area as the 802.11 network. Seamless roaming is achieved by “handing over” the call from one network type to another. In this work we will concentrate on handing over the ongoing call from the 802.11 network to the cellular infrastructure. A handover has to be scheduled. This is because the actual voice packets will not arrive at the handset for at least four seconds from once the handover was scheduled. There is a time lapse when establishing a call to the cellular or PSTN network, on average is it approximately five seconds to the GSM cellular network. Therefore we would have liked to predict the call quality this far in advance. Unlike the cellular system in which the network infras- tructure provides support for handovers, in an 802.11 system it is often the responsibility of the terminal to handover to a different access point. This it should do, on its own, by measuring the network and physical level parameters and deciding when a handover should be scheduled. Since users are often moving, predicting their quality in the future is not an easy problem. Also as one of the main incentives of having dual network access for the same user is monetary, we should allow more conservative or more aggressive handovers to be made. By conservative we mean switch from the 802.11 to the cellular network early, hence reducing the risk of poor quality of even disconnection. On the other hand a more aggressive policy might permit remaining in the 802.11 coverage for a longer time, hence saving money. This of course assumes that the 802.11 access is cheaper than that provided by the cellular network. Therefore handovers should be “tunable” for different usage scenarios as shown in figure 1. We will return to the issue of prediction and determining the handover point in section IV. Monetary cost Conservative phone Dual-mode Aggressive strategy Handover points Less time in 802.11 Poor quality risk is lower More time in 802.11 Poor quality risk is greater Monetary savings strategy Fig. 1. Users can choose a more conservative or aggressive handover strategy. The handover decision is based on real-time network mea- surements. The goal of the measurements is to estimate the real-time quality of the call as perceived by the listener. Rather than solely base our estimation on measuring network parameters, we also used PESQ an ITU-T standard, to estimate a number of consecutive lost packets that would correspond to a certain quality reduction [5]. Therefore we can at least attribute losses, approximately, to the quality that would be experienced by the listener. Knowledge of the losses alone, however, do not yield sufficient knowledge in order to trigger a successful handover. We need to ascertain if the losses are temporary or will become increasingly worse as time proceeds. Many users are tolerant of temporary glitches, particularly those who use mobile phones regularly. However if they become too frequent, the tolerance decreases rapidly. In order to establish the cause of the disturbance it is useful to look at additional network parameters to see if the situation is actually deteriorating or is indeed temporary. We will consider two possible causes of impaired speech quality, poor network coverage or heavy traffic loads. In these situations it is conceivable to consider