From: Chuck Reeves (creeves@windows.microsoft.com)
Date: Sun Oct 17 2004 - 23:04:13 PDT
The paper, "Congestion Avoidance and Control" was written in 1988 by Van
Jacobson at University of California-Berkley. It describes a number of
modifications to TCP that it suggests could help alleviate radical drops
in network bandwidth due to congestion (called congestion collapse). The
bulk of the text is spent focusing on 3 aspects of the TCP stack
behavior. The first describes a self-clocking technique (called
slow-start) for establishing an appropriate rate of packet transmission
during connection initialization. It uses an adjustable packet window to
control the number of packets transmitted to a receiver before an
acknowledgement is received. The window size is based on the rate of
receipt of acknowledgements from the receiver. The second analyzes the
appropriate behavior of the sender when dealing with loaded receivers.
Specifically, it proposes exponential damping of the average
round-trip-time and in-turn the retransmit timer. The final section
characterizes adjustments to the window for endnodes in the face of
congestion (lost packets due to lack of buffer capacity).The adjustments
allow senders to dynamically adjust transmission rates linearly upward
when bandwidth is available and exponentially downward when a timeout is
experienced.
Most of the math in this paper is supported heuristically with only
modest statistical data and focuses exclusively on an isolated and
specific approach. I found this comment in section 2 particularly
indicative of the author's unwillingness to discuss other approaches to
solving some of these problems. "Only one scheme will work, exponential
backoff, but proving this is a bit involved". This pattern continues in
section 3, "Without justification, I'll state that the best increase
policy is to make small, constant changes to the window size:". It would
seem more effective for the author to present these ideas as hypothesis
and then back them up with statistical measurement.
This archive was generated by hypermail 2.1.6 : Sun Oct 17 2004 - 23:04:19 PDT