Has internet delay gotten better or worse?

Abstract

Delay is a key Internet performance metric and its stability, variation, and abrupt changes have been well studied. However, little could have been said about the Internet-wide delay distribution. In order to build a representative sample set for the Internet-wide delay distribution, one needs to draw data from a random selection of source hosts to destination hosts and there is no measurement system with access to every AS and subnet of the Internet. In this work we propose to apply the path-stitching algorithm to archival measurement data and reconstruct the past history of Internet delay distribution. The two main advantages of path stitching are that data from existing measurement projects is sufficient to provide accurate estimates and it produces delay estimates between almost any two hosts in the Internet. As a first step towards the longitudinal study of the Internet-wide delay distribution, we examine how the Internet delay changes from 2004 to 2009. Our work is the first ever systematic approach to Internet delay distribution. We report the overall delay distribution has gotten worse from 2004 to 2009, while the delay distribution for the same set of host pairs remains almost identical or slightly improved.