From: Danny Wyatt (danny@cs.washington.edu)
Date: Sun Oct 17 2004 - 21:45:44 PDT
Congestion Avoidance and Control
Van Jacobson
This paper presents solutions to 3 causes of network congestion: (1)
slow-start window growth for networks that get saturated before they can
reach equilibrium, (2) improved RTT estimation for maintaining
equilibrium once its reached, and (3) exponential backoff to control
congestion once it has occurred. I was very impressed with how
elegantly all three of these solutions fit together---both with each
other and with the existing network. For example, I was impressed by
the fact that linear slow-start, when considered in the larger, temporal
behavior of the network, ends up being exponential and thus nicely
complementing exponential backoff. And the two can be co-implemented in
4 lines of code using integer arithmetic! That's elegance if I've ever
seen it.
I was also pleased to see (what I've learned as) statistical learning
methods used for the improved RTT estimator. Since the RTT estimator
must be as accurate as possible to allow for timeouts to bear some
information about network congestion, this is a neat validation of those
techniques.
My only complaint about the paper is its somewhat bizarre organization.
It's put together as 3 papers in one: the high-level, speculative and
explanatory main text; the theoretical justifications for (some) of
those speculations in the appendices and footnotes; and the empirical
validation of the speculations in the figures. They have certainly
covered all of their bases, but I think they could have covered them in
a more organized presentation.
This archive was generated by hypermail 2.1.6 : Sun Oct 17 2004 - 21:45:45 PDT