Failure Analysis and Modeling in Large Multi-site Infrastructures

Every large multi-site infrastructure such as Grids and Clouds must implement fault-tolerance mechanisms and smart schedulers to enable continuous operation even when resource failures occur. Evaluating the efficiency of such mechanisms and schedulers requires representative failure models that are able to capture realistic properties of real world failure data. This paper shows that failures in multi-site infrastructures are far from being randomly distributed. We propose a failure model that captures features observed in real failure traces.