We consider the problem of allocating data center (DC) resources for cloud enterprise customers who require guaranteed services on demand. In particular, a request from an enterprise customer is mapped to a virtual network (VN) class that is allocated both bandwidth and compute resources by connecting it from an entry point of a data center to one or more hosts while there are multiple geographically distributed data centers to choose from. We take a dynamic traffic engineering approach over multiple time periods in which an energy-aware resource reservation model is solved at each review point. For the energy-aware resource reservation problem, we present a mixed-integer linear programming (MILP) formulation (for small-scale problems) and a heuristic approach (for large-scale problems). Our heuristic is fast for solving large-scale problems where the MILP problem becomes difficult to solve. Through a comprehensive set of studies, we found that a VN class with a low resource requirement has a low blocking even in heavy traffic, while the VN class with a high resource requirement faces a high service denial. Furthermore, the VN class having randomly distributed resource requirement has a high provisioning cost and blocking compared to the VN class having the same resource requirement for each request although the average resource requirement is same for both these VN classes. We also observe that our approach reduces the maximum energy consumption by about one-sixth at the low arrival rate to by about one-third at the highest arrival rate this also depends on how many different CPU frequency levels a server can run at. (C) 2017 Published by Elsevier B.V.
展开▼