Multi-megawatt datacenters, while critical to our economy, can pose a severe cost, scalability and environmental burden. Imposed/Voluntary demand response programs, to address these concerns, mandate dynamic power regulation of a datacenter. This requires each of its hundreds of thousands of servers to expose and manage knobs for dynamic regulation. As datacenters resort to increasing levels of application consolidation on high-end servers to reduce their Total Cost of Operation, the co-located applications contend for the finite resources at each server. While prior research has looked at solutions for alleviating this contention for conventional computing resources, co-located applications also contend for the dynamically varying power budget, which is becoming an increasingly scarce resource. This project intends to elevate power to become a first class citizen for explicit management, similar to other hardware resources, in the context of current and next generation servers hosting multiple applications.
This project will specifically investigate the following issues: (i) when should server power be regulated? (ii) what amount of the regulated power should be allocated to each application, and recursively to each of its computing resources? (iii) how should power allocations be enforced and isolated? The project will develop evaluation platforms to implement and validate the developed research ideas in each of these 'when', 'what' and 'how' dimensions, using real datacenter/cloud workloads. The research contributions from this project can transform the design of server hardware and software, which make their way into large datacenters. The project proposes curricular enhancements at undergraduate and graduate levels using this research, in addition to the involvement of women, minorities and undergraduate students.
|Effective start/end date||7/15/17 → 6/30/22|
- National Science Foundation: $500,000.00