Predicting system performance for multi-tenant database workloads

Database consolidation is gaining wide acceptance as a means to reduce the cost and complexity of managing database systems. However, this new trend poses many interesting challenges for understanding and predicting system performance. The consolidated databases in multi-tenant settings share resources and compete with each other for these resources. In this work we present an experimental study to highlight how these interactions can be fairly complex. We argue that individual database staging or workload profiling is not an adequate approach to understanding the performance of the consolidated system. Our initial investigations suggest that machine learning approaches that use monitored data to model the system can work well for important tasks.