Paper Review : End-to-End Packet Delay and Loss Behaviour in the Internet

Reviewer : Seh Leng Lim

The paper aims to model the packet delay and loss behaviour in the Internet. This will help to facilitate understanding of the packet delay and loss behaviour in the Internet which is important for designing network algorithms such as routing and flow control algorithms.

The main contribution of the paper is its estimate of Internet traffic using a simple single server queueing model with 2 input streams, where one stream is the probe traffic and the other stream is the Internet traffic.

The key main ideas are :

(a) Probe packets accumulate behind large Internet packets, also known as probe compression; probe compression becomes less frequent as the interval between 2 consecutive probe send increases

(b) Probe losses are random as long as probe traffic uses less than 10% of the available capacity

I think that the paper has some modest contribution (rating of 3) to the study of packet delay and loss behaviour in the Internet. The paper does attempt to provide a model whose analytical results correlate with experimental results.

The author’s analytical data does seem convincing especially when it reveals that the bottleneck is the transatlantic link between France and the United States. However, the author does not explain why he uses Lindley’s recurrence equation in his analysis to show probe compression. Also, the experiments are conducted with both the host and destination on the same machine to avoid time synchronization problems. This is somehow less realistic than when both host and destination are geographically distributed. The probe packet loss process is modeled using packet loss probability. However, I do not have a fair idea how the probabilities are computed from experimental data.

The author is able to prove that his model works with 2 links, namely the INRIA-UMD and the UMD-University of Pittsburg link. Even then it is not clear if it works with other links as well. There is also an abnormal behaviour whereby stationary loss probability is large when the interval between sending two consecutive probe packets increases. The author attempts unconvincingly to attribute this abnormal probe packet loss to the faulty interface cards in the Suranet which is crossed by the INRIA-UMD link. I am thinking why the author does not try to repeat the experiment with other links that do not cross Suranet.

Researchers and builders who build Internet applications, especially those requiring buffering such as audio and video applications, will have a better feel of the packet loss and delay behaviour in the Internet from this paper.