In-the-field Predictor performance

In addition to simulations, we have evaluated the performance of the different predictors also by using the implemented service prototype, built on top of our middleware, and by moving 4 laptop clients among the campus wireless localities during streaming provisioning. Even if the number of considered in-the-field handovers is largely lower than the simulated one (and, thus, less relevant from the statistical point of view), in-the-field performance results confirm the simulation-based ones. The prototype has achieved great average results for ABS, SH%, and ABD, also due to the lower number of APs considered, and the consequent simplification of correct handover prediction.

Table 3 presents predictor performance indicators collected with Orinoco Gold and Cisco Aironet 350 PCMCIA wireless cards either on a Windows XP SP2 and a Linux Fedora Core 2 laptop.

Orinoco Gold wireless card performs SP handover strategy. Regardless the operating system either ABS, SH% and ABD shows great performances, validating our handover predictor.

Cisco Aironet 350 wireless card performs HP handover strategy when ACU (Aironet Client Utility) Manager imposes "Periodically Scan for a Better Access Point". However it is not possible to perform AP scanning through NDISUIO driver when ACU Manager is turn on. In order to perform AP scanning it is mandatory to impose "Use Another Application to Configure My Wireless Settings"; in this case handover strategy is reactive. This is the reason because in table 3 we define Reactive Cisco 350 card handover strategy. Also in this case of reactive handover strategy our predictor shows great performance; moreover again GM-filtered based predictor performs better than actual-RSSI based one!
Minor comment: Cisco card on Linux OS does not return RSSI in dB; we exploit a Cisco table to transform RSSI in dB. (for details see "You Believe You Understand What You Think I Said..." by Joshua Bardwell)

Wireless Card	HO strategy	Operating System	Exploited RSSI	ABS (KB)	SH%	ABD (s)	pre-Ho buffer fulfill	post-Ho buffer fulfill
Orinoco Gold	SP	Windows XP SP2	Actual	145	80	5.01	193	9
			GM-filtered	131	100	6.97	199	12
		Linux FC2	Actual	109	100	4.70	199	13
			GM-filtered	100	100	8.07	199	12
Cisco Aironet 350	Reactive	Windows XP SP2	Actual	172	100	4.22	199	13
			GM-filtered	169	100	5.99	199	12
		Linux FC2	Actual	165	100	4.79	199	12
			GM-filtered	160	100	7.13	199	13

Table 3. In-the-field performance indicators.
Predictor parameters may differ from simulation to in In-The-Field experiments; see Table 4 for further details.

Finally let us stress occasionally we have experienced a significant degradation of performance indicators in the case of extreme RSSI fluctuations, e.g., when a client follows a trajectory in strict proximity of relevant obstacles, such as the reinforced concrete walls of our campus buildings. In those case we claim handover prediction is not possible.