Unsupervised Learning of Dynamic Resource Provisioning Policies for Cloud-Hosted Multitier Web Applications

Dynamic resource provisioning for Web applications allows for low operational costs while meeting service-level objectives (SLOs). However, the complexity of multitier Web applications makes it difficult to automatically provision resources for each tier without human supervision. In this paper, we introduce unsupervised machine learning methods to dynamically provision multitier Web applications, while observing user-defined performance goals. The proposed technique operates in real time and uses learning techniques to identify workload patterns from access logs, reactively identifies bottlenecks for specific workload patterns, and dynamically builds resource allocation policies for each particular workload. We demonstrate the effectiveness of the proposed approach in several experiments using synthetic workloads on the Amazon Elastic Compute Cloud (EC2) and compare it with industry-standard rule-based autoscale strategies. Our results show that the proposed techniques would enable cloud infrastructure providers or application owners to build systems that automatically manage multitier Web applications, while meeting SLOs, without any prior knowledge of the applications' resource utilization or workload patterns.