Power-efficient and Predictable Data Centers with Sliding Scheduled Tenant Requests

Power efficiency and predictable performance have become major concerns for cloud service providers as they significantly affect cloud adoption and tenancy cost. Providing guaranteed resources for predictable performance in data centers drives the need for a request model which abstracts the traffic characteristics as well as the resource requirements of tenant applications. In this paper, we propose a novel Sliding Scheduled Tenant (SST) request model which enables tenants to request their resources for an estimated required time duration which can slide within a certain time-window. We investigate the power efficient resource-guaranteed Virtual Machine (VM)-placement and routing problem for dynamically arriving SST requests. The problem requires provisioning of the specified resources in a data center for the required duration of requests by choosing an appropriate start- and end-time within their specified time window, so as to maximize the number of accepted requests while consuming as low power as possible. We develop a mixed integer linear programming (MILP) optimization problem formulation based on the multi-component utilization-based power model. Since this problem which is a combination of VM-placement, scheduling and routing problems, is computationally prohibitive, we develop a fast and scalable heuristic algorithm. We demonstrate the effectiveness of the proposed algorithm and SST request model in terms of power saving and acceptance ratio through comprehensive simulation results.