Dynamic Cloud provisioning for scientific Grid workflows

Scientific computing requires an ever-increasing number of resources to deliver results for growing problem sizes in a reasonable timeframe. In the last decade, while the largest research projects were able to afford expensive supercomputers, others were forced to opt for cheaper resources such as commodity clusters or computational Grids. Today, Cloud computing proposes an alternative by which resources are no longer hosted by the scientists' computational facilities, but leased from specialized data centers only when and for how long they are needed. In this paper, we analyze the problem of dynamic provisioning of Cloud resources to scientific workflows that do not have sufficient Grid resources available, as required by their computational demands. We propose and study four provisioning aspects that deal with the general leasing model encountered in today's commercial Cloud environments based on resource bulks, fuzzy descriptions, and hourly payment intervals: Cloud start, instance type, Grid rescheduling, and Cloud stop. We study the impact of our techniques to the overall execution time, overall cost, and cost per unit of saved time with respect to various instance types offered by the Amazon EC2.