Gentlefolk,
I invite all interested parties to review the results of an experiment
we recently completed at AppliedTheory Communications. The experiment
tests the validity of using BGP AS Path Lengths as a basic heuristic of
'network distance' when trying to determine which geographically
distributed mirror is optimal. The obvious advantages to this strategy
are:
* The passive nature introduces no additional traffic into the network
* End to End Semantics are preserved
* Clients incur no extra RTT latency as part of decision making
A large group of approximately 20,000 servers is studied with respect
to this algorithm, as are a number of smaller groups that reflect real
life mirrored networks. Factors such as mean RTT, RTT variance, and
packet loss are all considered. The viability of using Path Lengths as
a discriminatory criteria in light of their somewhat un-even
distribution is also discussed.
The results provide a stronger than anticipated indication that this
is an effective strategy to proximate cluster selection within
mirrored resource environments.
The study (in a variety of formats) and a reference implementation of
a path length server (which requires a read-only internal BGP v4 feed)
are available at:
http://pat.appliedtheory.com/bgp-prox/
Questions, Comments, etc to me. (mcmanus@AppliedTheory.com)
Best Regards,
Patrick
This archive was generated by hypermail 2b29 : Thu Nov 18 2004 - 11:21:25 MST