STI: Effective Diagnostic Strategies for Wide Area Networks

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

A critical missing piece of the end-to-end network performance puzzle stems from the fact that current network diagnostic strategies do not adequately take into account the effects of path delay. This is a proposal to develop extensions to existing diagnostic tools which will take path delay into consideration effectively, compensate for a variety of delay times, and test the effects of these new diagnostic tools with network users and operators using actual high performance applications. End-to-end Internet diagnosis is such a difficult problem because the symptoms of nearly all flaws scale with path delay and the symptoms that scale with delay cause classical diagnostic strategies to yield misleading results. Classical diagnostic tools and strategies yield confusing results because merely changing the path delay changes the symptoms of the flaws, even though the flaws are not in the additional part of the path. This new diagnostic methodology uses several different techniques to compensate for the effects of path delay on diagnostic results: Emulated Delay, which uses specialized software to buffer and delay packets as though they traverse a long path, Parametric Model Scaling, a method of verifying that a short path can support TCP bulk transport over a long path by verifying that the short path implements the required properties for TCP to meet the end-to-end performance objective, and "Scenic" Virtual Private Networks, where packets can be delayed by using various IP tunnels or virtual private networks to send them on long "scenic" alternate routes. These techniques will be used in conjunction with existing diagnostic tools and strategies. This project will develop a suite of enhanced diagnostic tools that compensate for the effects of path delay. This suite will be targeted at several different audiences: true end users, application developers, network administrators and diagnostic experts or tool developers. The suite will be built from a small kit of core components that implement the previously described techniques that compensate for path delay, in conjunction with some simple diagnostic applications. A prototype diagnostic infrastructure consisting of diagnostic servers and test targets placed in the network will be deployed. The prototype diagnostic infrastructure will be deployed in and around the Pittsburgh and the Front Range GigaPoPs, selected directly connected campuses, as well as at a few other key sites near collaborators. The prototype diagnostic infrastructure will be used to validate our diagnostic suite by recruiting real end users, application designers and network administrators to field test our tools. This effort will be patterned after the "early adopter" effort of the Web100 project. In summary, there are a vast number of performance problems in the greater Internet that remain hidden because they do not exhibit any symptoms when tested on short, low delay, paths. The goal here is to revolutionize Internet performance diagnosis by properly compensating for delay and eliminating false positive diagnostic results for short paths, resulting in effective diagnostic strategies for Wide Area Networks.