SPECTRAL BASED METHODS THAT STREAMLINE THE SEARCH FOR FAILURE SCENARIOS IN LARGE-SCALE DISTRIBUTED SYSTEMS

摘要

We report our work on the development of analytical and numerical methods that enable the detection of failure scenarios in distributed grid computing, cloud computing and other large scale systems. The spectral (i.e. eigenvalue and eigenvector) properties of the matrices associated with a non-homogeneous absorbing Markov Chain are used to quickly compute the long time proportion of tasks completed at a given setting of parameters. This enables the discovery of critical ranges of parameter values where system performance deteriorates and fails.