From: Ethan Katz-Bassett (ethan@cs.washington.edu)
Date: Sun Oct 31 2004 - 23:01:38 PST
In this paper, the authors present a revised metric to replace the delay
metric used in ARPANET routing decisions at that time. The earlier approach
led to oscillations between alternate routes. The old system used measured
delay as a predictor of link delay, but this ignores the effect of many
flows simultaneously switching routes based on this prediction. The authors
do a great job presenting the drawbacks of that metric, explaining it in
detail and offering a simple clear example.
As an alternative, the authors present HN-SPF (hop normalized). This metric
performs like the previous one under lighter loads, but betters it under
heavier loads. High delay links should not necessarily be avoided by all
flows. Under heavy load, the new metric works toward giving good
performance to average routes. It attempts to move routes that have
slightly longer alternatives.
One key feature of HN-SPF is that it does not affect the underlying SPF
algorithm used to find routes; it simply changes the metric used to weight
each edge. This fact made HN fit easily into the existing scheme and made
adoption simpler; however, it also limited what the authors are able to do.
The authors give a simple, adoptable solution to a flaw in the routing
mechanism. However, their decisions about the maximum value for a line and
for the relative weighting of different lines were, apparently, based on a
"value judgment." It seems like they could have justified this better or
compared different alternatives.
This archive was generated by hypermail 2.1.6 : Sun Oct 31 2004 - 23:01:53 PST