Dynamic workload patterns prediction for proactive auto-scaling of web applications

Abstract

Proactive auto-scaling methods dynamically manage the resources for an application according to the current and future load predictions to preserve the desired performance at a reduced cost. However, auto-scaling web applications remain challenging mainly due to dynamic workload intensity and characteristics which are difficult to predict. Most existing methods mainly predict the request arrival rate which only partially captures the workload characteristics and the changing system dynamics that influence the resource needs. This may lead to inappropriate resource provisioning decisions. In this paper, we address these challenges by proposing a framework for prediction of dynamic workload patterns as follows. First, we use an unsupervised learning method to analyze the web application access logs to discover URI (Uniform Resource Identifier) space partitions based on the response time and the document size features. Then for each application URI, we compute its distribution across these partitions based on historical access logs to accurately capture the workload characteristics compared to just representing the workload using the request arrival rate. These URI distributions are then used to compute the Probabilistic Workload Pattern (PWP), which is a probability vector describing the overall distribution of incoming requests across URI partitions. Finally, the identified workload patterns for a specific number of last time intervals are used to predict the workload pattern of the next interval. The latter is used for future resource demand prediction and proactive auto-scaling to dynamically control the provisioning of resources. The framework is implemented and experimentally evaluated using historical access logs of three real web applications, each with increasing, decreasing, periodic, and randomly varying arrival rate behaviors. Results show that the proposed solution yields significantly more accurate predictions of workload patterns and resource demands of web applications compared to existing approaches. ? 2018 Elsevier Ltd