PROV-TE: A Provenance-Driven Diagnostic Framework for Task Eviction in Data Centers

Cloud Computing allows users to control substantial computing power for complex data processing, generating huge and complex data. However, the virtual resources requested by users are rarely utilized to their full capacities. To mitigate this, providers often perform over-commitment to maximize profit, which can result in node overloading and consequent task eviction. This paper presents a novel framework that mines the huge and growing historical usage data generated by Cloud data centers to identify the causes of overloads. Provenance modelling is applied to add contextual meaning to the data, and the PROV-TE diagnostic framework provides algorithms to efficiently identify the causality of task eviction. Using simulation to reflect real world scenarios, our results demonstrate a precision and recall of the diagnostic algorithms of 83% and 90% respectively. This demonstrates a high level of accuracy of the identification of causes.