SPX: SOPHIA: Support for Opportunistic Parallelism with Heterogeneous Intermittently-powered Accelerators

物联网的兴起依赖于越来越多的移动的和众多的计算设备，对于这些计算设备，基于电池的或有线的操作并不总是可行的，并且对于这些计算设备，能量收集（例如，太阳能）操作可能是优选的。越来越重要的工业、社会和政府领域将从采用能量收集或混合动力解决方案中获益，例如健康自省、智能城市、增强现实传感、辅助网络物理系统、安全和国防监控、应急响应，以及与绿色计算的自然协同作用。预计能量收集设备数量众多，但不可靠，因此在将其用作并行计算基础设施时既提供了机会，也带来了挑战。该项目开发了从软件到硬件的新接口，允许将配备片上非易失性存储器的能量收集设备视为并行计算加速器，同时保持系统级可靠性。该项目通过培训研究生和本科荣誉生克服跨层挑战，将研究与教育结合起来，从电路到系统软件，利用经常发生电源中断的计算平台。该项目解决了使用能量收集动力系统的并行计算中的三个开放性挑战：1）硬件和软件之间在向前进展保证方面缺乏一致性和协同作用; 2）过度关注串行执行场景，这并不适合能量收集系统中的并行性;以及3）先前努力探索能量收集系统的建筑支持的设计空间的特设性质。该项目引入了四个新的基础抽象来解决这些开放的问题：1）一个统一的，跨层的硬件-软件接口，用于计量和管理向前的进度; 2）一组公平和并行的向前的进度度量; 3）一个公开的跨层接口，用于推理近似向前的进度;以及4）用于非易失性处理器（NVP）的基于卸载的模型，其中NVP可以被视为用于“机会计算”的加速器。“项目结果包括这些抽象的验证，硬件支持的近似和异构能量收集机会计算加速器的设计空间的系统映射，以及开发和公开分发新的软件工具和模拟基础设施，这些工具和基础设施是模拟在能源上执行的复杂软件系统所必需的-该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The rise of the Internet of Things relies on increasingly mobile and numerous computing devices, for which battery-based or wired operation is not always feasible and for which energy-harvesting (e.g. solar power) operation may be preferable. There are increasingly important industrial, social, and governmental domains that will profit from adopting energy harvesting or hybrid power solutions, such as health introspection, smart cities, augmented reality sensing, assistive cyber-physical systems, security and defense monitoring, and emergency response, as well as natural synergies with green-computing. Energy-harvesting devices are expected to be numerous, but unreliable, and thus offer both an opportunity and challenge in their use as a parallel computing infrastructure. This project develops new interfaces spanning from software through hardware that allow energy-harvesting devices equipped with on-chip nonvolatile memory to be treated as parallel computing accelerators, while maintaining system level reliability. This project integrates research with education by training graduate students and undergraduate honors students in overcoming the cross-layer challenges, from circuits through system software, in utilizing computation platforms that experience frequent power interruptions.This project addresses three open challenges in parallel computing using energy-harvesting powered systems: 1) A lack of coherence and synergy between hardware and software in terms of forward progress guarantees; 2) An overfocus on serial execution scenarios that has not lent itself to parallelism in energy-harvesting systems; and 3) the ad-hoc nature of previous efforts to explore the design space of architectural support for energy-harvesting systems. The project introduces four new foundational abstractions to address these open questions: 1) an unified, cross-layer hardware-software interface for metering and managing forward progress; 2) a set of fair and parallel forward progress metrics; 3) an exposed cross-layer interface for reasoning about approximate forward progress; and 4) an offload-based model for nonvolatile processors (NVPs), wherein NVPs can be viewed as accelerators for "opportunistic computation." Project results include the validation of these abstractions, a systematic mapping of the design space of hardware-supported approximate and heterogeneous energy-harvesting opportunistic computation accelerators, and the development and public distribution of the new software tools and simulation infrastructure necessary to model complex software systems executing on energy-harvesting accelerators.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.