Generating and Executing Dependable Application Software on UnReliable Embedded Systems (Get-SURE) - II

随着技术的不断发展，计算机系统中的硬件部件变得越来越不可靠，由软错误（如瞬态故障）和永久错误（如老化）引起的硬件错误已经不可避免。因此，系统设计者利用有效的硬件和软件技术来提高可靠性是至关重要的。本研究计划的目标是在考虑不可靠硬件组件的情况下，采用跨层的方式采用可靠性驱动编译、离线和在线软件技术来实现可靠性的提高。为了弥合硬件和软件技术之间的差距，以实现高度的可靠性，该项目将探索软件弹性建模（考虑适当的硬件级故障模型）和优化，以指导系统软件和应用软件生成（通过编译器）实现可靠性改进。在降低了第一阶段（2013-2014年）的错误概率之后，本课题将在第二阶段（2015-2016年）研究在过程变化和老化效应下，具有协同多粒度错误检测和恢复的可靠性指导编译和系统软件技术。特别是，我们将探讨：(1)考虑具有异构硬化级别的核心的可靠性/弹性，应用程序和执行时间模型；(2)老化影响下的可靠性驱动编译器后端和链接器；(3)编译驱动的多粒度错误检测与恢复模式设计与优化；(4)基于功率/性能约束的多核系统可靠性优化；(5)编译器引导的运行时支持错误检测和恢复。离线和在线技术都是必需的，并且将在项目中进行探索，以便找到如何使它们相互适应的有效手段。

As hardware components in computer systems have become more unreliable due to the technology scaling, hardware errors, resulting from soft errors (e.g., transient faults) and permanent errors (e.g., aging), have become unavoidable. It is, therefore, crucial for system designers to exploit effective hardware and software techniques to improve the dependability. This research proposal aims to achieve dependability improvement by adopting reliability-driven compilation, offline and online software techniques in a cross-layer fashion while considering unreliable hardware components. To bridge the gap between hardware and software techniques for achieving a high degree of reliability, the project will explore software resilience modeling (considering the proper hardware-level fault models) and optimization to guide system software and application software generation (through compiler) towards reliability improvement. After reducing the error probabilities in our funding phase-I (2013-2014), this proposal for our funding phase-II (2015-2016) investigates reliability-guided compilation and system software techniques with collaborative multi-granularity error detection and recovery under process variations and aging effects. In particular, we will explore: (1) reliability/resilience, application, and execution time models considering cores with heterogeneous hardening levels; (2) reliability-driven compiler back-end and linker under aging effects; (3) compiler-driven design and optimization of multi-granularity error detection and recovery modes; (4) power-/performance-constrained reliability optimization for manycore systems; and (5) compiler-guided run-time support for error detection and recovery. Both offline and online techniques are required and will be explored in the project in order to find effective means of how to adapt them to each other.