Profiling temporal event behavior for demand prediction in cloud application performance management

Abstract

To sustain a good viewing experience of Internet live event broadcast service for users, application performance management in the presence of highly dynamic and unpredictable demand relies on a close grasp of the demand behavior characteristics and an accurate prediction model of them. In this paper, we propose a learning-based behavior profiling model which takes event-related temporal information into account, and separately characterized and classified the demand behavior of event periods rather than for the entire event as a whole. We also propose a run-time prediction algorithm based on the generated demand characteristic profiles and the state transition probability matrix to support an accurate forecast of the external demand in dynamic resource allocation for target performance management. The results show that our proposed model can well capture the demand temporal dynamics and changes, as well as minimize the probability of target performance violation while making a good utilization of resources in the presence of an unpredictable and highly dynamic workload.