Paper
review: General AMID
Congestion Control
Reviewer:
Kevin Hofstra
- What is the best algorithm
for congestion avoidance? Should
all protocols, such as UDP and other best effort, time sensitive also be
forced to abide by congestion parameters?
Can flow parameters be altered to allow different sending rates?
- An evaluation of the different
congestion control algorithms, and an attempt to make improvements on
their efficiency, fairness, and convergence speed. Investigation into how parameters effect sending rates.
An analysis of the GAIMD as TCP-Friendly.
-
- The most effective
algorithm is GAIMD, which changes window size by a multiplicative factor
during a triple-duplicate ACK, changes it to 1 if there is a timeout,
and uses an additive increase if it is successful.
- The relationship
between alpha and beta for GAIMD is shown to be TCP-Friendly. The flow rate of
GAIMD flows have approximately the same sending rate as a TCP
flow under the same conditions.
- The backoff reduction of 7/8 decreases the rate fluctuation,
and conversion speed when compared with the old value of ½. By reducing the deviation from the
optimum flow rate, the convergence speed to maximum efficiency is
increased at a slight cost of larger packet lost in very high
congestion.
- You can control flow
parameters so that flows with different values can achieve different
sending rates.
- Critique the main
contribution
- Significance- 3
The article is very effective in showing how different
algorithms for congestion avoidance will reduce lost packets,
responsiveness, and convergence, but is more an investigation of how to
alter existing solutions to be more effective.
- Convincing- 4 The data gathered supports their proposed
solutions. They are able to show
that by controlling flow parameters they are able to achieve varying
sending rates. The GAIMD flow rate
is demonstrated to be TCP-Friendly, which is one of their major successes
in a heterogeneous environment.
- System researchers and
builders should recognize that the major point in the success of GAIMD is
the fact that it has been proved TCP-Friendly. Therefore it will not throw off the
balance that TCP congestion control is establishing. The fact that is is
true while modifying the parameters of backoff
allows GAIMD to be fine tuned for best performance. We have reduced the backoff
in attempt to increase efficiency and reduce conversion time, and the
paper shows us that this is possible without sacrificing congestion
avoidance stability.