From: Rosalia Tungaraza (rltungar@u.washington.edu)
Date: Tue Oct 19 2004 - 22:45:43 PDT
This paper is about congestion avoidance using the increase/decrease
algorithm. In short, congestion is monitored by routers, which in turn
expect all end-hosts to respond appropriately. When the load level on the
network is above the optimal level, the routers attach a bit along the
header sections of incoming packets that signal their respective sources
to reduce the number of packets they are sending. Each source gets this
information from ACK packets originating from the hosts to which they sent
the packets (destination host).
Even though I didn't understand them in their entirety, I think the
algebraic (mathematical) formulation of how to bring about congestion
avoidance is one of this paper's strengths. I especially favor the
diagrams that accompanied some of the equations/expressions because they
made those expressions more understandable.
Despite the fact that the mechanism proposed by the authors for congestion
avoidance is plausible (and has been implemented e.g. DECbit), I do wonder
about one aspect of their scheme. How do they define a user's primary
bottleneck resource? In other words, what if a user passes through several
routers, each with a high potential of being congested? Is it fair to only
give that user the same proportion as every other user even if one of the
other users only used one router for the entire travel-distance of its
packets?
I understand, that fairness is a hard concept to implement in
inter-networks because the definition is arguable. The algorithm in this
paper makes this situation even worse because it deliberately
decentralizes decision-making. Hence, the routers don't get much feedback
from user packets in terms of how many other routers they have already
traversed before reaching a given router. Nonetheless, I think this paper
could slightly improve if the idea of fairness in relation to my questions
was elaborated more.
In conclusion, the authors point out some questions/areas that could be
explored further. These include exploring how delayed feedback affect the
control, whether it is important to know the number of users in the
network, and what will happen if the number of bits were increased to more
than one for signaling to users about the current state of the network.
This archive was generated by hypermail 2.1.6 : Tue Oct 19 2004 - 22:45:43 PDT