Understanding scheduling implications for scientific applications in clouds

Abstract

This paper explores some of the effects that the paradigm of Cloud Computing has on schedulers when executing scientific applications. We present premises regarding to provisioning and architectural aspects of a Cloud infrastructure, that are not present in other environments, and which implications they may have on scheduling decisions in presence of relevant policies like improving performance. We also argue that using virtualization as a mechanism for workload consolidation in a multi-core environment has important performance consequences for e-science. We propose and test a preliminary workload classification, based on usage modes, that may improve early scheduling decisions as we research towards automatic deployment of scientific applications.