This paper focuses on the problem of using real-time interactive audio applications in a packetized Internet environment. The main mechanism used is to buffer the packets at the receiver end and begin playback only when just enough packets are arrived. If buffering is too small, then certain packets may arrive after their designated playback time, and they will be lost. If buffering is too large, then the latency may be too long, and this effect is intolerable for conversations. Because propagation delay and network delay are unknown, the trick is to estimate their values.
This paper provides, analyzes, implements, and compares four different algorithms that estimate the network delays, so that the correct length of buffering time can be calculated. Three of these four approaches are based on either previous studies in this area or ideas drawn from related topics, such as how to calculate the round-trip time. The fourth algorithm is one that is developed by the authors and will be elaborated on in the following paragraph.
All of the first three algorithms fail to address the issue of delay spikes, which are sudden increases in the network delay time of certain packets. These algorithms do not behave very well in these conditions. The fourth algorithm that the authors developed addresses this issue by employing a simple heuristic. The idea is that if the difference between the two network delay values of the two most recent packets is greater than a certain threshold, then a different formula is used to estimate the network delay. This formula uses only the most recently observed delay values to estimate future delays. A variable var is used to determine when to return to normal mode - when var becomes small enough. Results from this paper show that the fourth algorithm generally performs better than the other three when the variance in the delay is large.
I give this paper a rating of 3 - modest contribution. The authors address one problem with the three known algorithms and provide a new algorithm that solve the problem - the problem of spiked delays. I actually do not have much criticism against this paper. It solves a problem with a simple heuristic, which I like. (Simplicity is good.) I also like the fact that the authors implemented and tested their algorithms in a real network environment using part of the Internet rather than using a simulation. I do think that the authors should have provided the others graphs based on varied buffer size. Although I understand that space is a constraint, sometimes not providing all the data may make it seem like the authors are trying to hide something.