SHF: Large: Collaborative Research: Designing the Programmable Many-Core for Extreme Scale Computing

Purdue University, West Lafayette IN

Investigators

Abstract

This work proposes to design a programmable many-core for Extreme-Scale Computing in mobile platforms (netbooks and smart-phones) of year 2020. This work cuts across the architecture, compiler, operating system, and correctness/performance tools areas. A key technology explored is that of cores and all of the software continuously operating in Chunks (i.e., atomic blocks) of instructions at a time --- eliminating the need for in-order, single-instruction-at-a-time commit. The PIs will develop a novel chunk-based architecture that supports the high levels of performance, power/energy efficiency, concurrency, and locality required. They will develop advanced compiler support for chunk generation that delivers high performance at low power, and leverages all the programmability features of the architecture. They will also design an OS that supports and takes advantage of chunks. Finally, they will design a set of novel correctness and performance tools that exploit chunks, signatures, hashes, and all the other features of this architecture. The broader impacts of this work involve the creation of a multidisciplinary research and education center at University of Illinois and Purdue on Programmable Extreme Scale Computing. Faculty of diverse expertise will be devoted to solving the problem of programmable, very-high performance, very power/energy-efficient many-cores for mobile platforms of year 2020 and beyond. The PIs will broaden the course offerings at University of Illinois and Purdue in the four areas, with multidisciplinary courses at different depth levels. Graduate and undergraduate researchers in ECE and CS will be involved in the research. Overall, the PIs hope to prove that programmable, high-performance, and highly power/energy-efficient many-cores based on continuous atomic-block operation are attractive.