From: Chandrika Jayant (cjayant@cs.washington.edu)
Date: Sun Oct 17 2004 - 22:37:59 PDT
"Congestion Avoidance and Control"
Written by Van Jacobson
Reviewed by Chandrika Jayant
Three solutions for congestion avoidance and control are
provided in this paper: the slow-start method to reach equilibrium,
improved round trip timing variation and retransmit backoff methods, and
a congestion avoidance scheme(warn and fix). Van Jacobson clearly says
three ways that packet conservation can fail and addresses each of them
separately. Though treated as separate algorithms, he nicely integrates
them into the existing TCP system with very little code, and they can
work together as one system. I like how he makes connection equilibrium
akin to physics- it makes it seem very natural and easy to abstract and
improve. I also like how he speaks about the linearity of networks and
how exponential improvements are the only way to reach stability. This
is a very sound argument. He presents good solutions to deal with
increasing load size while not messing up smaller load size flows.
I have two main problems with this paper (other than the
barrage of typos!). The first is the many assumptions Van Jacobson
makes. He often states something big and then says it is too hard to
explain, but it is hard to just believe and build from that as a reader.
The other is the frequent mention of other papers either doing the same
thing, merely footnoting methods that are the creative ideas of someone
else. At times it seems like he is just putting together ideas that
already exist and making it more formal and cohesive- which is not a bad
thing and often necessary.
I also had more technical questions. Could there be a better
way to find the limit of the network rather than pushing the packets to
the limit? The window adjustment policy seems a bit iffy as he mentions
"handwaving" to find a decrease term and a packet increase that is
"almost certainly too large." I would have liked to see more about this.
The relevance of this paper is similar to many of the other
papers we have read- learning the motivations that built TCP/IP into
what it is today, and understanding the underlying framework. Van
Jacobson mentions that eventually one might desire a third term in the
Taylor series expansion of the load equation (a second order model),
when the Internet has "grown substantially." Since this was written in
1988 I wonder if that is used today. Future work mentioned deals with
extending the paper's ideas to gateway (as opposed to endpoint)
congestion detection and striving for more reliability.
This archive was generated by hypermail 2.1.6 : Sun Oct 17 2004 - 22:38:15 PDT