Managing data centers is a challenging endeavor. State-of-the-art management systems often rely on analytical modeling to assess the performance, availability, and/or energy implications of potential management decisions or system configurations. In this paper, we argue that actual experiments are cheaper, simpler, and more accurate than models for many management tasks. To support this claim, we built an infrastructure for experiment based management of virtualized data centers, called JustRunIt. The infrastructure creates a sandboxed environment in which experiments can be run on a very small number of machines using real workloads and real system state, but without affecting the on-line system. Automated management systems or the system administrator herself can leverage our infrastructure to perform management tasks on the on-line system. To evaluate the infrastructure, we apply it to two common tasks: server consolidation/expansion and evaluating hardware upgrades. Our evaluation demonstrates that JustRunIt can produce results realistically and transparently, and be nicely combined with automated management systems.
PDF
download
878987797
028-85220240
