Over the past 45 years, computer scientists and engineers have transformed plentiful transistors into computational capability. At present, energy costs jeopardize further capability scaling, especially in datacenters. Given constrained datacenter power budgets, which are already at megawatt scales, this project is developing new paradigms to design and deploy heterogeneous datacenter architectures to improve energy efficiency by an order of magnitude. Heterogeneity specializes hardware for particular applications and deploys a mix of hardware for a mix of application needs. Specialization improves efficiency as overheads from supporting general, arbitrary computation are reduced. To realize heterogeneous efficiency, this project is rethinking design and deployment. Rethinking design to reduce specialization costs, this project is parameterizing hardware designs to facilitate their re-use, applying statistical inference to explore design spaces, and clustering applications that have similar hardware needs to increase each design?s market volume. Rethinking deployment to reduce performance risk, this project is developing multi-agent systems that transparently manage and allocate heterogeneous resources to meet performance targets.
This project has the potential to impact the composition and management of datacenters that underpin cloud computing. As access to cloud computing is democratized, diverse users will seek computation. And this project is architecting energy-efficient infrastructure to provide that computation. The design and deployment of heterogeneous architectures requires a holistic view of computer systems. The project is integrating research with curricular development, broadening students? views of processor, memory, network, and storage. The project is engaging all students at all levels of education and from all backgrounds, preparing future generations for computer engineering and research.
