Scalable fault localization in enterprise networks

Dipu John John, Purdue University

Abstract

Modern enterprise networks encompass tens of thousands of network entities and present a very challenging task of monitoring the network. Because of the sheer size of these networks, fast, accurate, automated and scalable fault localization becomes one of the primary objectives of any enterprise network fault management system. Finding the root cause for hard failures or performance related failures in large scale networks is a difficult task since the dependencies among various network entities are generally complex, dynamic in nature and span multiple levels. In this work, we present a scalable fault localization technique called Spotlight, which constructs a multi-tier graph where all possible dependencies among the network entities are modeled as a graph, then approximates the dependency graph into a bi-partite graph and finally runs a greedy inference algorithm to localize the faults. Representing the network as a multi-tier graph enables us to capture the real multilevel dependencies that exist in the network; but performing inference on muti-tier graphs is not scalable since the graphs can grow really large even for moderately sized networks and the time to localize the fault increases exponentially. The major contributions of this thesis include an efficient technique for compressing a multi-tier graph into a bi-partite graph and a scalable, accurate algorithm to perform inference on the compressed graph. The possibility of compressing a multi-tier graph into a two-tier model without significant information loss presents great potential from the scalability aspect. We show that Spotlight achieves over 90% accuracy for most faults in enterprise networks and is at least 100 times faster than similar prior approaches.

Degree

M.S.E.C.E.

Advisors

Rao, Purdue University.

Subject Area

Electrical engineering|Computer science

COinS