CAREER: Addressing Scalability Challenges in Designing Next-generation GPU-Based Heterogeneous Architectures

图形处理单元（gpu）正在成为许多领域的默认加速器，例如高性能计算（HPC）、深度学习和虚拟/增强现实。它们与高性能多核CPU架构的紧密集成也允许非常高效的异构计算。展望未来，这种基于gpu的系统必须在性能和能源效率方面进行扩展，以满足未来的百亿亿级（甚至更高）计算需求。然而，这些系统的持续扩展是具有挑战性的，主要是因为a)制造单个大芯片的成品率非常低，使其成本过高，b)内存层次结构仍然是关键的性能和能源效率瓶颈，c)可编程性和应用程序可扩展性受到共享虚拟内存和多应用程序支持效率低下的阻碍。该项目旨在通过重新思考未来大规模基于gpu的系统的设计来解决这些可扩展性挑战。特别是，本研究项目围绕三个主要组成部分展开：a)核心（包括其组织）和整个内存层次结构的设计空间探索；b)通过新的协同缓存和调度技术识别并利用缓存局部性来开发数据移动优化技术；c)通过增强共享虚拟内存和多应用程序执行支持来提高大规模系统资源的资源利用率。所有三个研究组成部分将在新开发的综合评价基础设施上进行评价。这项研究的结果将被纳入新的和现有的本科和研究生课程。预计从这项研究中得出的见解将对基于gpu的计算产生长期的积极影响，从而使我们的日常生活更富有成效。

Graphics processing units (GPUs) are becoming default accelerators in many domains such as high-performance computing (HPC), deep learning, and virtual/augmented reality. Their close integration with high-performance multi-core CPU architectures is also allowing very efficient heterogeneous computing. Going forward, it is imperative that such GPU-based systems scale both in terms of performance and energy efficiency to meet the exascale (and beyond) computing demands of the future. However, sustained scaling of these systems is challenging primarily because a) fabricating a single large die provides very low yield, making it prohibitively expensive, b) memory hierarchy remains a critical performance and energy efficiency bottleneck, and c) programmability and application scalability is hindered by inefficiencies in the shared virtual memory and multi-application support.This project seeks to address these scalability challenges by rethinking the design of future large-scale GPU-based systems. In particular, this research project revolves around three major components: a) design space exploration of cores (including their organization) and the entire memory hierarchy, b) development of data movement optimization techniques by identifying and then exploiting cache locality via novel synergistic caching and scheduling techniques, and c) improving resource utilization of large-scale system resources by enhancing shared virtual memory and multi-application execution support. All three research components will be evaluated on a newly-developed comprehensive evaluation infrastructure. The findings of this research will be incorporated into new and existing undergraduate and graduate courses. It is expected that the insights resulting from this research would have a long-term positive impact on GPU-based computing, thereby making our daily lives more productive.