Efficient workload management in geographically distributed data centers leveraging autoregressive models

The opportunity of using Cloud resources on a pay-as-you-go basis and the availability of powerful data centers andhigh bandwidth connections are speeding up the success and popularity of Cloud systems, which is making on-demand computinga common practice for enterprises and scientific communities. The reasons for this success include natural business distribution,the need for high availability and disaster tolerance, the sheer size of their computational infrastructure, and/or the desire to provideuniform access times to the infrastructure from widely distributed client sites. Nevertheless, the expansion of large data centersis resulting in a huge rise of electrical power consumed by hardware facilities and cooling systems. The geographical distributionof data centers is becoming an opportunity: the variability of electricity prices, environmental conditions and client requests,both from site to site and with time, makes it possible to intelligently and dynamically (re)distribute the computational workloadand achieve as diverse business goals as: the reduction of costs, energy consumption and carbon emissions, the satisfaction ofperformance constraints, the adherence to Service Level Agreement established with users, etc. This paper proposes an approachthat helps to achieve the business goals established by the data center administrators. The workload distribution is driven by afitness function, evaluated for each data center, which weighs some key parameters related to business objectives, among which,the price of electricity, the carbon emission rate, the balance of load among the data centers etc. For example, the energy costs canbe reduced by using a "follow the moon" approach, e.g. by migrating the workload to data centers where the price of electricityis lower at that time. Our approach uses data about historical usage of the data centers and data about environmental conditions topredict, with the help of regressive models, the values of the parameters of the fitness function, and then to appropriately tune theweighs assigned to the parameters in accordance to the business goals. Preliminary experimental results, presented in this paper,show encouraging benefits.