SPX: Enabling Scalable Synchronizations for General Purpose GPUs

点击翻译按钮获取中文摘要

Today?s GPUs have been successfully used for general purpose processing including high-performance computing, machine learning, graph analytics, to name a few applications. However, general purpose computing has different complexity and characteristics from those of graphics processing. A crucial characteristic is the need for coordination and synchronization among parallel threads, which occurs more often in general purpose applications than in graphics rendering. However, unlike common CPU designs, current GPU designs are missing full-fledged support for synchronization primitives that are convenient for implementing transparent scalability and achieving high performance for many parallel applications.This project develops truly effective, efficient, easy-to-use and easy-to-implement synchronization mechanisms for GPGPU while overcoming major obstacles pertaining to the GPU architecture. The first thrust will identify GPGPU synchronization primitives that are deemed sufficient to handle a large set of applications with static and dynamic dependencies. The second thrust will provide solutions and methodologies on how to use the new primitives in the most efficient way to suit the characteristics of applications. The third thrust addresses the main challenges in implementing those primitives in efficiency, possible context switching overhead, and memory capacity limitations. The success in this project will help fuel the spread of accelerator architectures for high-performance computing, cloud and mobile systems. It will increase efficiency for improved performance and reduced energy/power consumption, leading to a greener IT industry. Trained students will become future taskforce to continue the revolution of extending computing into every realm of science, life, commerce and culture.