One of the main questions in cloud computing environments is how to efficiently distribute user requests or Virtual Machines (VMs) based on their resource needs over time. This question is also an important one when dealing with a cloud federation environment where rational cloud service providers are collaborating together by sharing customer requests. By considering intrinsic aspects of the cloud computing model one can propose request distribution methods that play on the strengths of this computing paradigm. In this paper we look at statistical multiplexing and server consolidation as such a strength and examine the use of the coefficient of variation and other related statistical metrics as objective functions which can be used in deciding on the request distribution mechanism. The complexity of using these objective functions is analyzed and heuristic methods which enable efficient request partitioning in a feasible time are presented & compared.

Introduction

Federated cloud computing environments have recently emerged as a trending topic in cloud computing. Here Cloud Service Providers (CSPs) collaborate by delegating some (or all) of their customers’ requests to other CSPs. This 5 is done due to various reasons, be it overloaded servers in the federating CSP (i.e. the CSP that delegates parts of its request load to the federated CSP), the need to adhere to customer Service Level Agreements (SLAs) in special circumstances where the cloud provider cannot guarantee quality attributes, etc. Out of the decisions that must be taken in order to operate in such a federated envi10 ronment, one of the most crucial is which requests to federate and how should this federation take place keeping in mind the CSPs currently participating in the federation. The answer to this question must be one which is efficient and fair for all participating CSPs and incentivizes them to partake in the federation mechanism. How we model and evaluate this based on various objective 15 functions is an important consideration in this area. In this paper we emphasize a request distribution mechanism that focuses on multitenancy as a key factor that enables the cloud computing paradigm, providing many of the benefits of this paradigm from a cloud service provider perspective. The federation structure used here can be seen in Figure 1. Cus20 tomer requests are given to a Federation Broker who distributes them between CSPs which are cooperating in a cloud federation. Such a broker must consider multiple criteria when distribution occurs, including those relating to performance, pricing, quality of service, etc. The criteria (and its related objective functions) which we examine is to ensure request partitioning is efficient with 25 regards to utilizing multitenancy. To this end, different objective functions will be considered, including those that impact statistical multiplexing, as we will show throughout the rest of this section.