Detecting behavioral variations in system resources of large data centers

The identification of significant changes in systemresource behaviors is mandatory for an efficient managementof data centers. As the dimension of modern data centersincreases, the evaluation of state change detections throughtraditional algorithms becomes computationally intractable.We propose a novel approach that characterizes the statisticalproperties of the resource measures coming from systemmonitors, classifies them, and signals a change only whenthere is modification of the resource classification. This methoddiminishes the computational complexity and reaches the samedetection accuracy of traditional approaches as demonstratedby several results obtained in real enterprise data centers.