| In the Internet network dynamics such as congestion and route changes can cause lost or delayed transmissions. Such faults can significantly affect application performance. Diagnosis of these faults within the network is difficult. The goal of this dissertation is to design and evaluate techniques that enable the diagnosis of application perceived faults. Our approach is based on approximately isolating the location of application perceived faults and subsequently probing in the isolated fault area in order to diagnose the fault. Our techniques are based on (a) careful and adaptive end-system probing, and (b) obtaining different perspectives by probing from distributed locations.; We successfully demonstrate our approach through the design and evaluation of distributed, probe based fault isolation techniques for IP multicast. Receivers in multicast sessions are instrumented with “multicast traceroute” to periodically collect loss statistics, delays and path information along the multicast distribution tree. The receivers co-ordinate with each other to share monitoring responsibilities and to isolate, perhaps approximately, faults affecting reception. We develop currently deployable as well as router assist based schemes. Router assist schemes are based on “Subcast” and “Directed Multicast”. We find that while only one deployable scheme based on “TTL-scoping” has somewhat acceptable behavior while the router assist based schemes have very desirable scaling characteristics in addition to fairly accurate fault isolation capability. A secondary focus of this dissertation is to design and evaluate techniques to collect topology information, an important component of fault diagnosis. While our probing techniques may help approximately isolate faults, topology information may be required to infer the exact cause of the fault. We designed and evaluated techniques to track the inter-domain topology of the Internet. These techniques, based on analyzing inter-domain routing information, obtained fairly accurate maps. We also developed distributed and self-configuring techniques to enable scalable extraction of router-level topologies annotated with information useful for fault diagnosis. |
