Collaborative Research: Software Canaries

随着晶体管变得越来越小，它们越来越容易受到其时序特性变化的影响。处理可能变化的一种方法是在硬件设计期间假定最坏情况特性。不幸的是，最坏情况设计的功率和性能成本正变得令人望而却步。因此，处理器设计者寻求更激进的设计风格，以允许优于最坏情况的操作。其中一种方法使用“金丝雀电路”来指示处理器何时由于变化而处于不安全操作的边缘。这项研究调查了“软件金丝雀”——运用处理器最易受影响的电路来确定一个激进的安全操作点的软件例程。软件金丝雀实现了硬件金丝雀电路的所有优点，而没有相应的硬件开销。更重要的是，它们可以提供额外的好处，因为它们可以跟踪老化和温度——这些变化对于硬件金丝雀电路来说可能是无法区分的。此外，与硬件金丝雀电路不同，它们允许在每个应用程序的基础上识别最激进的操作点。本研究的重点是探索软件金丝雀的好处，开发用例，开发为不同微体系结构构建软件金丝雀的技术，以及开发支持基于软件金丝雀的执行所需的微体系结构和系统支持。这项研究更广泛的影响在于，通过选择积极的操作点，使计算系统的能耗显著降低。其他具体的影响包括开源工具的开发，这些工具可以用于其他研究小组的进一步工作，以及在研究生和本科生课程中对变化和能源效率的更多讨论。

As transistors get smaller, they become increasingly susceptible to variations in their timing characteristics. One way to deal with possible variations is to assume worst-case characteristics during hardware design. Unfortunately, the power and performance costs of worst-case design are becoming prohibitive. Consequently, processor designers have sought more aggressive design styles that permit better-than-worst-case operation. One such approach uses "canary circuits" that indicate when a processor is on the verge of unsafe operation due to variations. This research investigates "software canaries" - software routines that exercise the most susceptible circuits of a processor to determine an aggressive safe operating point. Software canaries enable all the benefits of hardware canary circuits, without the corresponding hardware overhead. More importantly, they may provide additional benefits as they can track aging and temperature - variations that may be indistinguishable to hardware canary circuits. Furthermore, unlike hardware canary circuits, they allow identifying the most aggressive operating point on a per application basis. This research focuses on exploring the benefits of software canaries, developing use cases, developing techniques for building software canaries for different microarchitectures, and developing microarchitecture and system support needed to support software-canary-based execution.The broader impact of the research lies in enabling computing systems with significantly lower energy consumption by selecting aggressive operating points. Other specific impacts include development of open source tools that may be used for further work by other research groups, and a greater discussion of variations and energy efficiency in graduate and undergraduate coursework.