System Failure Forewarning Based on Workload Density Cluster Analysis

Each computer system contains design objectives for long-term usage, so the operator must conduct a continuous and accurate assessment of system performance in order to detect the potential factors that will degrade system performance. Condition indicators are the basic components of diagnosis. It is important to select feature vectors that meet the criteria in order to provide true accuracy and powerful diagnostic routines. Our goal is to indicate the actual system status according to the workload, and use clustering techniques to analyze the workload distribution density to build diagnostic templates. Such templates can be used for system failure forewarning. In the proposed system, we present an approach, based on workload density cluster analysis to automatically monitor the health of software systems and system failure forewarning. Our approach consists of tracking the workload density of metric clusters. We employ the statistical template model to automatically identify significant changes in cluster moving, therefore enabling robust fault detection. We observed two circumstances from the experiment results. First, under most normal status, the lowest accuracy value is approximate our theoretical minimum threshold of 84%. Such result implies a close correlation between our measured and real system status. Second, the command data used by the system could predict 90% of events announced, which reveals the prediction effectiveness of this proposed system. Although it is infeasible for the system to process the largest possible fault events in the deployment of resources, we could apply statistics to characterize the anomalous behaviors to understand the nature of emergencies and to test system service under such scenarios.