Use of statistical methods to reduce delays for media playback buffering

Receiver playback buffers are required to smooth network delay variations. A playback buffer operates by introducing additional delay, acting as a holding area for packets whose scheduled playback time is in the future. Packets which arrive after their scheduled playback time are considered late and are discarded. The algorithms commonly used in the Internet are reactive, in that measure immediate network delay variations and then adjust the buffer size to avoid lateness, resulting in potentially large end to end delays and frequent buffer changes. This paper works with an alternative predictive approach, which records historical information and uses it to make short-term predictions about network delay, with the aim of not reacting too quickly to short-lived variations. This allows an application-controlled trade-off of packet lateness against buffering delay, suitable for applications which demand low delay but can tolerate or conceal a small amount of late packets. This paper builds on previous work by proposing aging techniques to improve delay predictions, evaluating performance using trace-driven simulation.