Review: Quantifying the causes of Path Inflation

This paper discusses the root causes of network paths being significantly
longer than necessary. In order to quantify the causes of path inflation
author has used a trace driven study of 65 ISP with a focus on topology
and routing policies at different levels. The data set consists of 19
million traces collected using 42 measurement sources in PlanetLab. The
different layers studied include intra-domain routing, inter domain
routing and ISP peering using this study author has categorized the impact
of each routing level and routing policies used at these level on path
inflation.

In intra-domain topology author observes that most of the paths are not
inflated by much. Also the inflation amongst tier-1 ISPs are higher than
other ISPs mainly because tier-1 ISPs are geographically spread and thus
connecting all the points directly is not feasible. Also it was observed
that the routing policies within an ISP are governed by the latency
sensitive routing thus having little impact on inflation. Analyses also
reveal that the peering topology also seems to have little impact on path
inflation and this impact is least between networks with more peering
points. On the other hand the peering policies cause path inflation and
this is mainly because of early-exit policy used by different ISPs. The
same inter domain level in which it seems topology causes modest path
inflation while inter-domain routing policies causes significant path
inflations. The mains reason for higher path inflation for both peering
policies and inter-domain policies is the BGP. BGP does not provide enough
support to enable ISPs to communicate with each other to share their
routing policies. Although ISPs tries to optimize but these routing
policies but lack of mechanism to achieve this goal make it hard to
implement.

I believe this is a good paper that gives an idea of how different routing
policies affect the path latencies. Also it has presented mechanisms to
come up with different routing policies used at different level which I
believe is a significant contribution.
Received on Thu Nov 16 2006 - 10:43:45 EST