From: Ethan Phelps-Goodman (ethanpg@cs.washington.edu)
Date: Sun Oct 17 2004 - 22:46:09 PDT
Congestion Avoidance and Control
Jacobson
This paper presents modifications to the TCP protocol that arose out of
a series of network failures. The Internet at the times was suffering
from catastrophic failures when the network became congested, despite
congestion control mechanisms built into TCP. Their proposals are all
simple. First, they give a slow-start mechanism so that a new
connection will edge up towards the available bandwidth rather than
exceeding it at first go. Second, they give a more sophisticated
timeout mechanism so that timeouts are more likely to genuinely lost
packets. One interesting point is that they are able to estimate and
optimize a parameter of the network (round trip variance) that
originally was a fixed constant. Finally, they advocate the additive
increase/multiplicative decrease in window size that we use today.
The protocols in this paper are simple but effective. They justify
their algorithms with strong analytical backing from queuing theory and
linear systems theory, as well as giving trace results from real
networks. The author had the great benefit of seeing a large scale
system in action. The original TCP attempted to deal with all these
issues, but they couldn't reasonably have known how their design would
scale in practice. The paper also had the interesting quirk of having
more than half its content in footnotes and appendices.
This archive was generated by hypermail 2.1.6 : Sun Oct 17 2004 - 22:46:32 PDT