From: Masaharu Kobashi (mkbsh@cs.washington.edu)
Date: Sat Oct 23 2004 - 13:01:44 PDT
1. Main result of the paper
The paper proposes a new gateway queuing algorithm for controlling
congestion and achieving fairness. It is claimed that the algorithm
outperforms the conventional FCFS queuing algorithm in fair allocation
of bandwidth, lower delay, and protection from ill-behaved sources
under various conditions including different flow control schemes.
2. Strengths in this paper
The algorithm allocate the vital three quantities, bandwidth, promptness,
and buffer space separately. It is logical and the best way to achieve
each of the sub-goals optimally.
One of the most valuable properties of the algorithm is it is capable of
protecting well-behaved sources from ill-behaved sources in the
allocation
of bandwidth.
In comparison to Nagle's algorithm the proposed FQ is superior in that
it takes into account the packet size in the queuing decision.
The algorithm has an important ability to provide a lower delay to
lower throughput sources independent of the window sizes of the
flow controls.
3. Limitations and suggested improvements
Argument on the acceptable definition of fairness is weak or nonexistent
in the paper. Since fairness includes subjective elements and also
depends
on assumptions on the environment, participants etc., to make a
scientific
argument, they should have presented results for each of possible
multiple
fairness standards. It should not make an argument based on a single
standard.
Ratio of net data to the header lengths is not taken into account in the
proposed algorithm. For small messages, the header length takes far
greater
weight than the cases of big file transfers. Therefore, with the propose
algorithm, small telnet messages are penalized unfairly.
Similar to the definition of fairness, the paper is weak in arguing the
justification of deciding "user". It chose each conversation as user.
But it is still controversial. Therefore, they should present the result
of the simulation for each of possible definition of user.
The environments for the simulations are too simple. But this problem
seems to be perceived by the authors, since they state that in their
future
work they need to investigate more realistic load conditions on larger
networks.
4. Relevance today and future
There seems to be a problem of gradual deployment with the proposed
algorithm. For example, if the definition of user and/or queuing
algorithms
are different at different gateways in the transition period for
deployment of this algorithm, you cannot expect the intended achievement
of the new algorithm for the whole of the Internet. Sporadic existence of
FQ can be even harmful because, for example, different definition of
fairness, and/or different definition of user at different gateways can
make the whole system unfair by the most common sense of fairness.
The simulations are based on only two types of services, telnet and
ftp. But there has been an increasing number of very different services
of very different requirements such as video/sound streams. To be
justified
to be a good contender for the current and future algorithm, it has to be
capable of handling those services with a more up-to-date notion of
fairness.
This archive was generated by hypermail 2.1.6 : Sat Oct 23 2004 - 13:01:45 PDT