Reviewer: Charlie Wiseman
Date: 9-15-2005
How would you rate this paper, relative to others we have read? top 50%, but not top 25%
How would you rate your knowledge of the topic of this paper? novice
What problem or issue does the paper address? Why is it important?
This paper shows how certain factors lead to throughput differences between TCP and TFRC. Clearly, this is something networking researchers and network administrators should be aware of when looking at real performance measurements and when trying to design more reliable protocols.
What are the main contributions of the paper and why are they important?
The authors show, both through empirical evidence and theoretical analysis, that competing TFRC and TCP flows can have very different lose rates. In addition, it is shown that lower sending rate flows (TCP or TFRC) can have higher lose rates than higher sending rate flows. Also, they find a correlation between these different lose rates and the initial difference in throughput. These all lead to the general conclusion of the paper, which is that there a few things that influence the long-term throughput of TFRC as compared to TCP.
How significant are these contributions relative to previous work?
As is mentioned in the paper, other work had already shown it likely that TFRC throughput didn't quite line up with TCP. So, there are really no original ideas presented in the paper. Instead, the authors take this possible incongruity and show that it does indeed exist. That said, the paper does contribute significantly to the understanding of this particular problem.
Give detailed comments justifying your view of the paper.
Clearly, this paper doesn't add much breadth to our collective understanding, but it does add depth. In particular, no great new ideas are explicitly suggested for rate controlled flows, but TFRC is carefully analyzed to find the reasons it doesn't match long-term throughput with TCP. So, then, this paper seems best used in two places: an aid for designing new (equation-based congestion controlled) protocols, and a reference for anyone who is looking at detailed network information involving both TFRC and TCP flows. While considering this paper, I tried to put myself into one of those positions to evalute its actual usefulness.
First, a note about the introduction. There were many terms and symbols being thrown around throughout the introduction without being really explained. As such, it was very difficult to really understand the points being made. Perhaps that section should have switched with the definitions section, since re-reading the introduction after having terms defined and assumptions listed made much more since.
Assuming that their assumptions are valid, the results should be quite useful. One of the most welcome aspects of the work done here is that there are both solid analytical proofs and detailed empirical data to rely on when drawing any conclusions. In general, they first give the theoretical basis for some condition, then show that it does indeed exist through simulation. In this way, they successfully show that the TFRC throughput equation, loss event estimation, and retransmission timeout estimation are all responsible for the throughput differences in TCP and TFRC. What is even more important is that all of their results are laid out in such way that the could potentially be used by others to successfully find ways around the problems encountered with TFRC.
Perhaps my lack of knowledge in this area is the problem now, but one thing I'm not clear on is why it matters so much. Yes, it's good to know that such imbalances can exist. But, are the details really that important? TFRC isn't TCP, and thus shouldn't necessarily be expected to behave exactly the same as TCP. The TFRC RFC says 'It is reasonably fair when competing for bandwidth with TCP flows'. I would probably define 'reasonably fair' to include the differences found in this paper. Specifically, one of the three big points in the paper is that the TFRC equation is part of the reason why there is a throughput difference. Well, that equation is specifically noted as being slightly different from the TCP equation. So, of course it will produce some different results.
My lack of understanding the need for this paper naturally drops my overall opinion of it, but as mentioned above, I believe this paper could be very useful to the right people. Hence, my rating.