Paper review : Charge-Sensitive TCP and Rate Control in the Internet (LA00)
Reviewer : Hai Fang (hfang@acm.org)
- Goal
To provide a network-independent algorithm which optimizes the system rates and
the total user utility.
- Contribution
This paper extends Kelly et al.'s work, which is rate-based, to the class of
window-based (TCP/Vegas) algorithms.
- Main ideas
- The system optimization problem can be decomposed into two subproblems, network
and user problems, in order to get a feedback-insensitive solution.
- When users update their parameters by solving their own optimization problem,
at an equilibrium the system optimum is achieved.
- In the case of a single bottleneck, there exists a unique Nash equilibria
when we model the problem as a noncooperative game.
- Evaluation
- Significance rating: 1
Most of the work in this paper are based on several previous work, especially on
Kelly's seminal work. The formalization and the analysis of the problem are copies,
in some sense, of the original work under the window-based case. And the authors
do not provide enough arguments to show the advantages and/or feasibility of their
approach.
- Convincing rating
The authors create and analyze their model for the optimization problem with
considerable mathematical teniques. They also use Network Simulator (NS) to
simulate their results.
However, the authors do not test their algorithm on the Internet; so the feasibility
of the algorithm on the realistic scale is still a question. Also, their simulation
uses the similar utility functions Ui, which is not so convincing for the possiblity
of the various Internet users. It seems to me not clear that why they can suppose the
user's utility function as multiple of log x.
There're also too many assumptions in this paper. Actually, this is partly caused
by the assumptions made by previous works, such as Kelly's and the TCP/Vegas protocol.
Anyway, the authors do not tell us whether the model/algorithm is still practical
when it satisifies so many restrictions. Some of the assumptions are not trivial,
in my eyes, such as each user has only a fixed route and a single bottleneck in the
model.
- Limitation
This work is done in a relatively simple model, e.g., it's only designed for unicast.
The authors should also consider how to combine their work with the currently dominant
TCP protocols, such as TCP/Tahoe or TCP/Reno. Since they cannot expect TCP/Vegas
can be more prominent by their paper.
- Conclusion
Pricing mechanism for a distributed system is complex. Appropriately modelling
the system and proposing a resonable and feasible definition of fairness, hence
providing an effective (network-insensitive) algorithm needs a lot of work.
10/29/01