Predictive Path Switching Control for Improving the Quality of Service in Real-Time Applications

A congestion avoidance strategy based on the concept of predictive path switching control (PPSC) is proposed for improving the quality-of-service (QoS) in real-time applications, and in particular in Voice over Internet Protocol (VoIP) applications deployed over the public Internet. The proposed PPSC strategy requires the dynamic prediction of traffic congestion levels over all available network paths. This predictive information is then used as an input for a control strategy to select the best available path at each pre-determined time. A study using emulated network paths explores the impact of overall path loss rate and traffic delay signal frequency content on the proposed PPSC strategy. The study reveals that PPSC provides the best QoS improvement for real-time applications if the average over-all loss rate of the available paths is between 5% and 15%, and if the traffic delay signal contains mostly low frequencies. Linear and nonlinear dynamic predictors are developed and used along with actual network data collected from PlanetLab for demonstrating the effectiveness of the PPSC strategy. The results show that PPSC is better than no path switching, with no one dynamic predictor providing the best performance for all case studies. A voting-based control strategy is proposed to overcome this problem. The results demonstrate that the voting-based control results in universally better performance for all cases studied. VoIP packets encoded with Speex, a publicly available encoder, demonstrate that any PPSC strategy is far more effective than no path switching. The proposed voting-based PPSC is moderately more effective than PPSC based on a simple predictor, both resulting in voice quality mostly over 3.0 in MOS, if the available network paths have overall loss rate in the range of 5% and 15%. Though encouraging, the PPSC strategy raises a number of implementation issues, which must still be addressed. In particular, scaling of the PPSC to a large number of