Extracting a Credible Hint of Response Time to Scale Resources in Elastic Clusters

Abstract

In elastic clusters, the service resources (or called “resources” for short) can be dynamically scaled, thus providing opportunities to cut down the energy cost of redundant resources. Generally, taking into account the Quality of Service (QoS) of clusters, resources are carefully scaled according to specific hints which are some features of system status. However, when the Service Quality Requirement (SQR) is referred to the response time of requests, some widely used features cannot well reflect the status of the QoS. Consequently, the QoS cannot be well maintained, and the energy-saving efficiency is unsatisfactory. In this paper, we indicate that under such SQR, the outstanding hint for resource scaling is the response time of requests. Accordingly, we propose a resource scaling method which scales resources leveraging an elaborate Hint of Response time (HR). More specifically, HR is credible to foresee future QoS, and our method extracts HR by tracking and making analysis on the waiting requests in each server. Moreover, when resource scaling operation is performed, our method can estimate how many resources are suitable for current workloads with a good accuracy. Thereby, our method can timely and directly scale resources to the suitable amount, thus can significantly reduce the time delay of re-matching resources. Finally, our method can significantly promote cluster performance on both the QoS and the energy-saving efficiency.