CNS Core Small: Simultaneous and Heterogeneous Multithreading

University Of California-Riverside, Riverside CA

Investigators

Abstract

The significance of artificial intelligence (AI), machine learning (ML), data mining, and virtual/augmented/mixed reality (VR/AR/MR) in our everyday life has led to the adoption of hardware accelerators in modern computers. The existence of heterogeneous processing models in a computer is now ubiquitous. However, the entrenched computing and programming approaches and frameworks prevent the concurrent use of hardware accelerators. As a result, most hardware circuits are idle when executing applications, wasting power, consuming energy, and reducing battery life but making no contributions to computation. This proposed project, simultaneous and heterogeneous multithreading (SHMT), will revisit the entire stack of system architectures, including memory systems, operating systems, programming languages, and applications, to enable the concurrent use of all available hardware resources. SHMT will improve the utilization of precious hardware resources, reduce the energy consumption of computational tasks, shorten the execution time of applications, and lead to better user experiences on modern accelerator-rich computer systems. The experience of developing SHMT will help innovate educational materials in related computer engineering courses and create research opportunities for under-represented minority-serving institutions where we perform main research activities. Enabling simultaneous execution of parallel computation on heterogeneous processors requires SHMT to tackle challenges in allowing the programming framework to generate code segments where each performs parallel partitions of calculation for a task, extending the runtime system to ensure the qualities of execution results despite the various precisions, and hardware characteristics of each computing resource, providing mechanisms and abstractions for data sharing and exchanging in minimizing the overhead from the divergent data formatting demands on multiple types of processing resources, and potentially optimizing the scheduling policies and mechanisms to provide higher flexibility in task allocation. Specifically, the proposed project contains the following major tasks. (1) A programming framework and a set of applications support the SHMT model. (2) An intelligent runtime system guarantees the quality of results. (3) An innovative memory subsystem facilitates concurrent task execution on diverse computing resources. And potentially (4) A runtime system maximizes efficiency. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.