This project develops energy management techniques for large-scale, multi-tier data-centers, using formal optimization methods that do not compromise performance requirements. High-performance Internet services are often functionally decomposed into a pipeline of multiple, semi-independent tiers spanning multiple machines. These in turn might consist of a heterogeneous mix of system generations, due to phased upgrades. Power delivery and cooling costs constitute a significant part of total cost of ownership.
The expected workload of such a system is aperiodic, with highly varying execution demand and possibly significant I/O activity. This represents an opportunity for major energy savings, but only if performance guarantees can be maintained. Scalability also requires that local power management decisions must yield globally optimal energy consumption. Although effective power management policies have been devised for simpler architectures, no straightforward extensions exist to make them applicable to multi-tier clusters while preserving response-time guarantees.
Results will be disseminated at leading conferences and the software will also be publicly available on the Internet for use by the research community in further research, or by the industry as a basis for possible product development. Moreover, from a global societal perspective, successful energy management helps save natural resources, while reduced service-provider costs foster economic growth and IT innovation. Finally, this research will help train graduate and undergraduate students.
