Network Troubleshooting from End-Hosts

Troubleshooting faults or performance disruptions in today's Internet is at best frustrating. When users experience performance or connectivity problems, there is little they can do. The most common attempt to solve the problem is to reboot or call the provider's hot line. Network providers have more information to use in diagnosing problems in their networks, but their network troubleshooting is often mostly manual and ad-hoc. We argue that automatically troubleshooting network faults or performance disruptions requires monitoring capabilities deployed at end-hosts (in or close to the customer's premises). End-host monitoring is necessary to detect the problems that affect end-users. In addition, when problems happen outside the control of the network administrator or the end-user performing the troubleshooting, end-host monitoring is the only approach to identify problem location. The goal of our research is to make network troubleshooting more transparent by designing tools that require as little as possible human involvement (both end-users and administrators). This document presents our initial steps towards automatic network troubleshooting from end-hosts as well as our long-term objectives. Our main contributions are the design of more accurate and efficient end-host measurement methods. We work both on techniques to detect faults and performance disruptions, and to identify the location of the problem. For fault identification, we improve the two basic techniques using end-to-end measurements: traceroute and network tomography. First, we show that traceroute, the most widely used diagnosis tool, reports erroneous paths in the presence of routers that perform load balancing. We build Paris traceroute to correct these errors. Second, we design measurement methods for accurate fault identification using network tomography. Network tomography assumes up-to-date and correlated measurements of the status of end-to-end paths and of the network topology, which are hard to get in practice. We design techniques to track correlated path reachability and network topology. Finally, we design techniques for lightweight detection of faults and performance disruptions. We minimize the overhead of active end-to-end probing for fault detection and design a tool to collect passive measurements at end-hosts