SHF: Small: Enabling and Analyzing Accuracy-aware Reliable GPU Computing

图形处理单元（gpu）正在成为通用硬件加速的默认选择，因为它们能够为大规模高性能计算应用程序提供数量级的速度和节能执行。由于在大规模HPC系统上执行的大多数此类应用程序都是长时间运行的，因此它们应对各种基于硬件和软件的故障非常重要。许多先前的研究表明，真正的高性能计算系统容易受到软错误的影响。缺乏必要的保护和检查点机制可能导致科学生产力降低、操作效率降低，甚至造成经济损失。然而，这些保护机制（例如，纠错码）本身并不是免费的——它们会产生非常高的性能、能源和面积成本。该项目采用了一种全面的方法来探索减少这些保护开销的途径，方法是利用所有错误都不会导致应用程序输出的准确性出现不可接受的损失这一事实。先前的结果表明，GPGPU应用程序可以适应这种精度感知优化。为了实现这些优化，该项目将解决三个主要研究问题：a)需要哪些硬件/软件支持和工具来确定哪些指令不容易受到软错误的影响；b)基于此分析，哪些硬件组件不需要保护，保护多长时间，同时不牺牲超出用户质量要求的应用程序质量；c)在资源管理和调度方面需要哪些优化，以使低开销但可靠的计算更有效和高效。这些问题将通过高性能计算（HPC）、大数据分析、机器学习和图形等领域出现的各种GPGPU应用来探讨。如果成功，该项目将产生一些新颖的研究见解，这将在实现低成本可靠的GPU计算方面发挥重要作用。这个项目的成果将纳入现有的和新的关于计算机结构和可靠性的本科和研究生课程，这将有助于培训学生，包括妇女和来自不同背景和少数群体的学生。

Graphics Processing Units (GPUs) are becoming the default choice for general-purpose hardware acceleration because of their ability to enable orders of magnitude faster and energy-efficient execution for large-scale high-performance computing applications. Since the majority of such applications executing on large-scale HPC systems are long-running, it is very important that they cope with a variety of hardware- and software-based faults. Many prior works have shown that real HPC systems are vulnerable to soft errors. An absence of essential protection and checkpointing mechanisms can lead to lower scientific productivity, operational efficiency, and even monetary loss. However, these protection mechanisms (e.g., error correction codes) are themselves not free -- they incur very high performance, energy, and area costs. This project takes a holistic approach to explore the avenues to reduce these protection overheads by taking advantage of the fact that all errors do not lead to an unacceptable loss in the accuracy of application output. Prior results show that GPGPU applications are amenable to such accuracy-aware optimizations. In order to enable these optimizations, this project will address three major research questions: a) What hardware/software support and tools are necessary to determine which instructions are not vulnerable to soft errors, b) Based on this analysis, which hardware component(s) need not be protected and for how long, while not sacrificing application quality beyond the user's quality requirements, and c) What optimizations in terms of resource management and scheduling are necessary to make low-overhead but reliable computation more effective and efficient. These questions will be explored via a variety of GPGPU applications emerging from the areas of high-performance computing (HPC), big-data analytics, machine learning, and graphics. If successful, this project will generate several novel research insights that will play an important role in enabling low-cost reliable GPU computing. The results of this project will be integrated into the existing and new undergraduate and graduate courses on computer architecture and reliability, which will facilitate in training students, including women and students from diverse backgrounds and minority groups.