arXiv:1309.2444v3 [cs.DC] 6 Nov 2013 Distributed Coalition Formation in Energy-Aware Cloud Federations: A Game-Theoretic Approach (Extended Version) Marco Guazzone [1] , Cosimo Anglano [2] , Roberto Aringhieri [1] , Matteo Sereno [1] [1] Department of Computer Science, University of Torino, Italy [2] Department of Science and Technological Innovation, University ofl Piemonte Orientale, Italy ABSTRACT Federations among sets of Cloud Providers (CPs), whereby a set of CPs agree to mutually use their own resources to run the VMs of other CPs, are considered a promising solution to the problem of reducing the energy cost. In this paper, we address the problem of federation forma- tion for a set of CPs, whose solution is necessary to exploit the potential of cloud federations for the reduction of the energy bill. We devise a distributed algorithm, based on cooperative game theory, that allows a set of CPs to coop- eratively set up their federations in such a way that their in- dividual profit is increased with respect to the case in which they work in isolation, and we show that, by using our algo- rithm, they are able to self-organize into Nash-stable federa- tions and adapt them to environmental changes. Numerical results are presented to demonstrate the effectiveness of the proposed algorithm. Categories and Subject Descriptors C.2.4 [Computer-Communication Networks]: Distributed Systems; K.6.4 [Management of Computing and Infor- mation Systems]: System Management General Terms Management,Performance Keywords Cloud Federation, Cooperative Game Theory, Coalition For- mation 1. INTRODUCTION Many modern Internet services are implemented as cloud applications consisting of a set of Virtual Machines (VMs) that are allocated and run on a physical computing infras- tructure managed by a virtualization platform (e.g., Xen [1], VMware [2], etc.). These infrastructure are typically owned by a Cloud Provider (CP) (e.g., Amazon AWS, Rackspace, Windows Azure, etc.), and are located into a (set of possibly distributed) data center(s). One of the key issues that must be faced by a CP is the reduction of its energy cost, that represents a large fraction of the total cost of ownership for physical computing infras- tructures [3]. This cost is mainly due to the consumption of the physical resources that must be switched on to run the workload. To reduce energy consumption, two techniques are there- fore possible for a CP: (a) to minimize the number of hosts that are switched on by maximizing the number of VMs allocated on each physical resource (using suitable resource management techniques [4, 5]), and (b) to use resources that consume less energy. Cloud federations [6], whereby a set of CPs agree to mutu- ally use their own resources to run the VMs of other CPs, are considered to be a promising solution for the reduction of energy costs [7] as they ease the application of both tech- niques. As a matter of fact, while each individual CP is bound to its specific energy provider and to the physical infrastruc- ture it owns, a set of federated CPs may enable the usage of more flexible energy management strategies that, by suit- ably relocating the workload towards CPs that pay less for the energy, or that have more energy-efficient resources, may reduce the energy bill for each one of them. In order to exploit the energy saving potential of cloud fed- erations, it is however necessary to address the question con- cerning its formation. As a matter of fact, it is unreasonable to assume that a CP unconditionally joins a federation re- gardless of the benefits it receives, while it is reasonable to expect that it joins a federation only if this brings it a ben- efit. In this paper, we address the problem of federation forma- tion for a set of CPs, and we devise an algorithm that allows these CPs to decide whether to federate or not on the basis of the profit they receive for doing so. In our approach, each CP pays for the energy consumed by each VM, whether it belongs to its own workload or to the one of another CP, but receives a payoff (computed as discussed later) for doing so. The algorithm we propose is based on cooperative game the-