Towards deterministic network diagnosis

Internet fault diagnosis is important to end users, overlay network service providers (like Akamai), and Internet service providers (ISPs). For example, with Internet fault diagnosis tools, users can choose more reliable ISPs. However, The modern Internet is heterogeneous and largely unregulated, which renders the Internet diagnosis an increasingly challenging problem. Though several router-based Internet diagnosis tools have been proposed [1, 2], these tools generally need special support from routers. For example, Tulip [1] requires the routers to support continuous IP-ID for generated ICMP packets. Also these ICMP-based tools are subject to ICMP rate limiting and are sensitive to crosstraffic. In contrast, many recently-developed tools for Internet Tomography use signal processing and statistical approaches to infer link level properties [3, 4, 5, 6] or shared congestion [7] based on end-to-end measurements of IP routing paths. The advantage of statistical inference is their fine diagnosis granularity, i.e., they infer the property of each virtual link (a consecutive subpath of an IP path with no branches [5]). However, the accuracy of the predicted link properties is subject to uncertainty in the model assumptions. Since true multicast does not exist in the Internet, many mulitcast-based tomography approaches [3, 4] have to use unicast for approximation, such as [8] and [9]. Thus the accuracy of the probe measurements heavily depends on the cross traffic in the network, and there is no guarantee of their accuracy.