Paper Review:
Dynamic Behavior of Slowly-Responsive Congestion Control Algorithms
Reviewer: Robert Dugas
Problem
The problem dealt with is basically whether current TCP-compatible congestion
control algorithms are ready for deployment with respect to their dynamic response
characteristics.
Contribution
Primarily, this paper determines the dynamic response characteristics of a
number of TCP-compatible congestion control algorithms and determines that
most are in fact pretty good.
Main Ideas
- Classifies variety of current research into groups based on their relation to TCP.
- Develops dynamic testing benchmarks and assiciated metrics for congestion control schemes.
- Concludes that most current TCP-compatible congestion control algorithms will work with
TCP in a dynamic environment, some pending the addition of self-clocking.
- Notes that slow-response algorithms often lose bandwidth to TCP in a highly dynamic environment.
Critique
Significance:3
This paper breaks little new ground in terms of originality or reshaping major ideas about
networking. However, it does serve to validate a large amound of pre-existing work.
Methodology:
The experiments and framework set forth in this paper likely constitute is most worthwhile
contribution. The authors succeed in establishing general benchmarks and then evaluation techniques
for the congestion control problem. In addition, they identify metrics such as convergence time,
responsiveness, and smoothness as valid yardsticks. The experimentation takes place using the
standard barbell topology and includes the RED protocol.
Limitation:
Although the tests presented do seem to provide usefull benchmarks, the ultimate
determination of deployment-ready compatibility should take place on a large scale network of
framework rather than a few computers at MIT.
Lessons:
Basically, we learn that most current congestion control algorithms will deploy
successfully with TCP, although most slow-response algorithms lose out bandwidth-wise in a highly
dyanmic environement. Also, the addition of self-clocking to certain slow-response algorithms
greatly enhances performance.