CS268 Reading
Review
Analysis of the Increase and Decrease Algorithms for Congestion
Avoidance in Computer Networks
Dah-Ming Chiu and Raj Jain
Review by Feng Zhou
1/29/2003
The problem: The central part of congestion avoidance algorithm is the
way to increase and descrease the load in face of congestion and normal
transmission. Linear control methods (additive and multiplicative) are studied
in this paper analytically.
Key points:
- The most significant conclusion is decrease must be multiplicative to ensure
the system to converge to a fair stable state. This not-so-obvious conclusion
is very clear in the vector representation graph used in the paper, which in my
opinion is a pretty useful tool. Beside that, the increase policy must have an
additive item and have a multiplicative factor >= 1.
- The second useful point is that in order for hosts to be able to make
decisions without global knowledge. The increase policy should be additive, in
the abscence of truncation. This, plus the first conclusion, leads to the AIMD
rule that is widely in use now.
- Regarding optimal convergence to effectiveness and fairness. One useful
result is by increasing "responsiveness", you always decrease "smoothness", with
is natural for such a feed-back control system. For fairness, using additive
increase always gives us fastest convergence.
- Although the conclusion is pretty reasonable and useful. The analysis
process seems overly simplied to me. Basically the simplest case in which two
hosts using the same policy is analyzed. However, in more complicated cases
like many hosts with different parameters, what will happen will be interesting.
And just as the authors pointed out, as the feedback is delayed and all hosts
are acting asynchronously rather than synchronously, how these factors impact
the convergence, efficiency and fairness of the system are still unsolved
problems.