Paper review : End-to-End Packet Delay and Loss Behavior in the Internet (Bol93)

• Goal
  To understand the packet delay and loss behavior in the Internet.

• Contribution
  This paper presents a simple method to measure the round trip delay and loss behavior in the Internet. Understanding this behavior is important for the proper design of network algorithms such as routing and flow control algorithms, for the dimensioning of buffers and link capacity, and for choosing parameters in simulation and analytic studies.

• Main ideas
  • The UDP echo tool can be used for measuring the end-to-end delay and loss behavior of the Internet.
  • The experiment results can be interpreted using a simple single server queueing model with two input streams, which represent probe traffic and Internet traffic respectively.
  • The observations not only agree with results obtained by others using simulation and experimental approaches, but also show interesting and less expected behavior, e.g., the randomness of the probe packets loss.

• Evaluation
  • Significance rating: 3
    Many studies of packet delay and loss in various network environments have been reported in the literate. (2)
    The method proposed in this paper is simple but effective in the scale of the whole Internet. (+1)

  • Convincing rating
    First, the author give a review on different approaches, including analytic, simulation, and experimental ones, for the delay and loss measurement problem. Then a simple experimental method using UDP echo tool was proposed. The experimental data was analyzed by a queueing model. The examples, e.g., the calculation of the bandwidth between France and US, the phenomenon of probe compression, support the correctness of the model.

  • Limitation
    This method needs a well-controlled experiment, say, the packets need to be send at regular intervals. It also need to know more information about the parameters and testing route than the general experimental methods.

• Conclusion
  Simple tool and simple model sometimes lead to a concise but useful analysis, if there is some.

• Remarks
  • It seems that there's a typo in the last column of Table 3, the ulp for delta=500ms should be 0.07 or 0.097 instead of 0.97 from the analysis followed the table.
  • For the bandwidth problem, is it feasible to do a test on each hop of the route from the source to the destination? For example, we may calculate the bandwidth between node 1 and i (1 < i <= 10) in Table 1, and get the bandwidth is b >> 128kb/s between node 1 and j (1 < j <5), whereas approximately 128kb/s between node 1 and k (5 <= k <= 10); hence we know that the bottleneck in this route is between node 4 and 5.

  9/18/01