AIMD fallacies and shortcomings
Prasad
1
AIMD claims:
Guess What !?
“Proposition 3. For both feasibility and optimal convergence to fairness, the increase policy should be additive and the decrease policy should be multiplicative.”
AIMD claim is untrue !
Consider the following simple example:
No. of users = 2
Init loads of users X1 = 17 and X2 = 0
Load goal, Xgoal = 20
Fairness goal, Fgoal = 99%
AIMD equations
Let aI = 1,aD = 0, bD = 0.01 and as per AIMD claim, bI should be 1
Fairness index is given by:
After plugging in all the values…
Result is (after 3 iterations):
Now, change bI to 1.1. In other words,
introduce a multiplicative-component during
increase. Result then is (after 3 iterations):
2
With AIMD, there is a possibility of unlimited overload after convergence
AIMD equations
After summing the values for n users we get,
Defining overload to be:
We getOverload =
The problem is, as n becomes large, overload becomes large as well !
3
AIMD is rather slow w.r.t convergence of efficiency
4
All issues mentioned till now have one thing in common – they are all related to the synchronous communication system
This model is too simple and unrealistic and hence, inferences made based on it may not hold at all in a real system
And Guess what !?
5This is the best part !
AIMD does not guarantee fairness !
(in a more realistic asynchronous communication system like the Internet)
A better model