CSR: Small: Efficient Many-core Execution Models for Cognitive Computing

University Of Connecticut, Storrs CT

Investigators

Abstract

Society is increasingly dependent on digital machines due to miniaturization and seamless integration of mobile, cloud and network computing. This project's objective is to develop an understanding of the computational algorithms, and devise novel architectural methods that improve the efficiency of the devices and platforms that will execute cognitive computing problems. Achieving good performance and scalability for concurrent hardware has been widely acknowledged to be a hard and important problem. This project singles out communication as a grand challenge and proposes novel methods for systems ranging from single chips with many "cores" or processors, through tightly integrated and increasingly heterogeneous futuristic many-core machines. The proposed research agenda provides for significant broader impacts related to (1) curriculum development and student training through industry collaborations and integration of cognitive computing in computer architecture curriculum, and (2) outreach through established Research Experiences for Undergraduates (REU) sites and supplement programs. A set of auxiliary communication models are proposed to give the programmer new mechanisms based on hardware-supported explicit messaging to exploit fine-grained parallelism in the algorithms that represent cognitive computing problems. The resulting architecture and software concurrency choices are expected to expose many problem-algorithm-input-machine configurations, and solving for a near-optimal configuration in real-time is a hard problem. A novel situationally adaptive execution model is proposed that analyzes and captures this massive search space to pick the right choices in the spatiotemporally changing runtime settings. The incorporation of auxiliary communication models in futuristic multicore and tightly-coupled many-core processors is timely as it will enable performance scaling trends to continue at ultra-energy efficiency: enterprises and US mission agencies will be able to deploy multicore technology in real-time embedded systems, such as self-driving cars. As a broader impact, better solutions for emerging cognitive computing problems will help improve our cyber infrastructure, healthcare, and manufacturing, all with significant benefits to the US economy and society.