Dynamic Bayesian network based prediction of performance parameters in cloud computing

Resource prediction is an important task in the cloud computing environment. It can become more effective and practical for large Cloud Service Providers (CSPs) with a deeper understanding of their Virtual Machines (VM) workload's key characteristics. Resource prediction is also influenced by several factors including (but not constrained to) data centre resources, types of user applications (workloads), network delay and bandwidth. Given the increasing number of users for the cloud, if these factors can be accurately measured and predicted, improvements in resource prediction could be even greater. Existing prediction models have not explored on capturing the complex and uncertain (dynamic) relationships between these factors due to the stochastic nature of the cloud systems. Further, they are based on score-based Bayesian network (BN) algorithms having limited prediction accuracy when dependency exists between multiple variables. This work considers time-dependent factors on the performance prediction of the cloud. It considers an application of Dynamic Bayesian Network (DBN) as an alternative model for dynamic prediction of cloud performance by extending the static capability of a Bayesian network (BN). The developed model is trained using standard datasets from Microsoft Azure (MA) and Google Compute Engine (GCE). It is found to be effective in predicting the application workloads and its resource requirements with an enhanced accuracy compared to existing models. Further, it leads to better decision making process with regard to response time and scalability in dynamic situations of cloud environment.