ARPANET

This paper focuses on the performance issues relating to routing metric
used in ARPANET. Author highlights that currently developed routing
metric is a strict delay metric i.e. the routing decisions are not based
on the relative network load, and because of this it performs well under
the average load but in a highly loaded environments it led to routing
instabilities, wasted link and processor bandwidth. The author suggests
that instead of strict routing policy the goal of a routing metric under
heavy load should be to give the average route a good path instead of
attempting to give all routes the best path.

The author then purposed a revised metric that acts similar to the
strict delay metric under light loads and switches to capacity based
metric under heavy loads to give average route a good path. One of the
problems with previous metric was of routing oscillation resulting
because of routing decisions that were not based on the relative link
cost. In the revised metric the author instead of simply relying only on
the delay cost, relies on normalization of delay cost by dividing it by
the cost of alternatives links, thus taking into account the effect of
cost relative to other links. The new metric also imposes the upper and
lower bounds on the cost values reported by lines to avoid unfair use of
network links. The author has demonstrated that once these changes were
implemented in ARPANET they resulted in better load sharing and general
performance and the metric is stable under most conditions.

The new metric is most effective when network traffic consists of small
node to node flows and this is because single path routing algorithms
are fairly ineffective in dealing with large flows. Further enhancements
could be made but that could require extensive change to the packet
forwarding system as well as packet header.
Received on Tue Sep 19 2006 - 08:02:15 EDT