Summary: Internet Path Inflation

The paper presents an analysis of root causes of Internet path inflation
namely topology and routing policy. Specifically, the paper examines how
internet path inflation occurs at three levels – Intra-domain (source and
destination that belong to the same network), peering factors (source and
destination belong to adjacent ISP) and Inter domain (source to destination
involves multiple ISPs).

To study intra-domain factors, the paper assumes that routing is weighted
shortest path with unknown set of edge weights. It outlines a constraint
based approach to infer these weights. The approach is based on the fact
that weight of an observed path must be less than or equal to that of any
alternative path between pair of nodes. The measurements indicate that
Intra-domain traffic engineering is a common place and has minimum impact on
path inflation.

Next, the paper studies inflation caused by topology and policy of peering
factors. Peering topology does not inflate paths significantly. However,
peering policies often cause path inflation and can sometimes lead to highly
inflated paths. This difficulty stems from the fact that BGP doesn't offer a
convenient way to do optimal path selection or topology sensitive load
balancing across peering links. Inflation could be decreased quickly by
increasing the number of peering points. Also, optimal exit peering policy
could be employed to achieve latency reduction.

The paper further studies the Inter-domain topology and policy factors. The
results indicate that Inter-domain routing causes significant amount of path
inflation with more than half of paths longer than the shortest path. Here
also, the main culprit the design of BGP. BGP doesn't allow for propagation
of enough information to enable an informed path selection. Consequently,
ISPs often end using minimum AS-hop count metric for routing which leads to
path inflation.

Roughly half of all inflated paths are due to inter-domain routing. This
inflation arises not from commercial constraints but from using AS-path
length as a routing metric. Finally, the observed inflation in paths in
peering and Inter-domain policies signifies that ISP lack good tools to find
better paths. ISPs should use right set of tools and optimize their own
paths for latency as long as it doesn't cause congestion.

I thought that paper is very well presented. Especially, I find that the
conclusions of the paper are well backed up solid measurement data.
Received on Thu Nov 16 2006 - 05:00:53 EST