SHF: Small: Collaborative Research: Resilient Computing Systems Using Deep Learning Techniques

SHF：小型：协作研究：使用深度学习技术的弹性计算系统

• 批准号: 1528045
• 负责人: Vijay Janapa Reddi
• 金额: $ 26.5万
• 依托单位: University of Texas at Austin
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2015
• 资助国家: 美国
• 起止时间: 2015-08-01 至 2018-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1528045&HistoricalAwards=false
• 关键词: SHF; Small; Collaborative; Research; Resilient

Over the past decade, computer systems have become prone to a variety of hardware failures. Traditionally, hardware failures were circumvented by operating the system at less than peak computing efficiency, effectively compromising efficiency to achieve reliability. Such a conservative approach is no longer a viable option because it leads to significant energy inefficiency. Since datacenters containing thousands of computers are one of the largest and fastest growing consumers of electricity, it is important to decouple the relationship between hardware failures and energy efficiency. The PIs' research will lay the groundwork for an intelligent computing system that operates at peak efficiency, but manages its fault resiliency and reliability using machine-learning based deep learning techniques. In effect, the system learns to steer itself clear of danger whenever its deep neural nets anticipate a failure. The research will address several important issues involving the scalability, flexibility and efficiency of deep learning techniques for various types of hardware failures. If successful, the research product will minimize, if not eliminate, penalties to the system that stem from the various circuit and micro-architectural techniques that are commonly used to mitigate and overcome hardware failures.

在过去的十年中，计算机系统已容易发生多种硬件故障。传统上，通过以低于峰值计算效率操作系统，可以避免硬件故障，从而有效地损害了实现可靠性的效率。这种保守的方法不再是可行的选择，因为它会导致效率不大。由于包含数千台计算机的数据中心是电力最大，增长最快的消费者之一，因此将硬件故障与能源效率之间的关系解除很重要。 PIS的研究将为以峰值效率运行的智能计算系统奠定基础，但使用基于机器学习的深度学习技术来管理其故障弹性和可靠性。实际上，每当其深层神经网预测失败时，该系统就会学会避免危险。该研究将解决涉及各种硬件故障的深度学习技术的可扩展性，灵活性和效率的几个重要问题。如果成功的话，研究产品将最大程度地减少源于各种电路和微构造技术的惩罚，这些系统通常用于减轻和克服硬件故障。