CSE 770 Paper Review

Reviewer: Michela Becchi
Date: 9-8-2005

How would you rate this paper, relative to others we have read? top 25%, but not top 10%

How would you rate your knowledge of the topic of this paper? familiar, but not expert

What problem or issue does the paper address? Why is it important?

The paper aims at proposing an accurate probe process for measuring end-to-end packet losses on network links. Specifically, the accuracy of the proposed model is determined in terms of loss frequency and duration.

Since packet losses are at the basis of several versions of TCP, the understanding of their characteristics is fundamental. On the other hand, the determination of good mechanisms for their measurement is still a field under study.

What are the main contributions of the paper and why are they important?

• Highlighting of the limitations of Poisson-modulated probes in packet losses’ measuring. Specifically, this is done in two ways: (1) theoretically, by relating the asymptotic nature of such methods with the long time period required in the measurement; (2) experimentally, by testing one Poisson-probe based tool on three different setups.
• Introduction of an alternative probe process in two distinct variant
• Validation and Evaluation of the proposed probe process

How significant are these contributions relative to previous work?

I believe that the significant contribution of this paper consists in the introduction of the new probe process. The introduction of a process not belonging to the typically used stochastic processes is something which seems innovative and promising.

Give detailed comments justifying your view of the paper.

After having discussed and empirically demonstrated the limitations of Poisson process based probes, the authors propose an alternate probe process for measuring network losses.

I believe that the paper is well structured. The discussion on the limitations of traditional probes is convincing. The new model is extensively described first in theoretical terms, and then validated through experimentation. The experimental setups are detailed, and the interpretations of the formulas are normally detailed. The assumptions made are also highlighted.

It is not clear to me how the estimator of the average loss duration in the extended variant of the algorithm corrects the bias for finite N, since R appears also in the quotient. My impression is more that this second version allows weaker assumptions than that it does correct bias, but my understanding can be wrong. It was quite difficult for me to clearly understand all the assumptions in the model and their motivations.

It’s also difficult for me to say whether the results are somehow biased by the experimental setup, and whether somehow ZING is somehow penalized in the comparison. I think that the choice of three different setups makes the discussion convincing.

Overall, the paper seems to me very good.