Negative impact of path switching

Potential Negative Impact of Path Switching

Although path switching can potentially improve application quality, there is also the possibility for introducing some performance degradation. For instance, in addition to the overhead incurred by path probing and quality estimation, switching from one path to another may also cause transient disruptions, especially when the candidate paths differ significantly in their propagation delay as this will result in an abrupt change in end-to-end delay. We argue that given the relatively low frequency of path switching decisions and the expected range of delay increase or decrease, those differences can be readily absorbed by most playback mechanisms.

Assume a voice flow that is switched from path A to path B, and where the sender transmits packets at a constant rate of r packets per unit of time and is unaware of the switch. Consider first the case where the end-to-end delay on path A is smaller than on path B. The switch then causes a reception ``gap'' at the receiver, which could potentially lead to discontinuity in voice playout at the receiver. Conversely, when the delay on path A is larger than on path B, the receiver will get a burst of out-of-order packets.

However, the use of a playout buffer together with adaptive playout control can mitigate negative effects in both above scenarios. Specifically, when path switching causes a sudden decrease of the buffer filling level, the receiver can adaptively reduce the decoding rate to avoid buffer underflow. Conversely, if path switching causes a sudden increase of the buffer filling level, the receiver can speed up the decoding rate to maintain the playout delay close to the optimal value. Since voice streams are typically composed of interleaved talk and silence spurts, such adaptation can be easily implemented by prolonging or shortening the silence spurt following the talk spurt affected by path switching. Experiments show that the negative impact of path switching, with adaptive playout control, is negligible when the delay difference between the candidate paths is lower than 100 ms. For example, below is a set of speech samples we tested in different path-switching scenarios.

911-noswitching.wav: a 55-second speech sample without experiencing path switching and playout adaptation.
911-switching-80ms.wav: the same speech sample experiencing path switching. Path switching is performed every 5 seconds (in a practical path switching system, the switching frequency is much lower) between two paths that have 80 ms propagation delay difference.
911-switching-200ms.wav: the same speech sample experiencing path switching. Path switching is performed every 5 seconds between two paths that have 200 ms propagation delay difference.
911-switching-1000ms.wav: the same speech sample experiencing path switching. Path switching is performed every 5 seconds between two paths that have 1000 ms propagation delay difference (unlikely in a real network setting, for demonstration only).