Self-economy in cloud data centers: Statistical assignment and migration of virtual machines

Abstract

The success of Cloud computing has led to the establishment of large data centers to serve the increasing need for on-demand computational power, but data centers consume a huge amount of electrical power. The problem can be alleviated by mapping virtual machines, VMs, which run client applications, on as few servers as possible, so that some servers with low traffic can be put in low consuming sleep modes. This paper presents a new approach for the adaptive assignment of VMs to servers and their dynamic migration, with a twofold goal: reduce the energy consumption and meet the Service Level Agreements established with users. The approach, based on ant-inspired algorithms, founds on statistical processes: the mapping and migration of VMs are driven by Bernoulli trials whose success probability depends on the utilization of single servers. Experiments highlight the two main advantages with respect to the state of the art: the approach is self-organizing and mostly decentralized, since each server locally decides whether or not a new VM can be served, and the migration process is continuous and adaptive, thus avoiding the need for the simultaneous reassignment of many VMs. © 2011 Springer-Verlag.