Enabling adaptation in embedded and real-time systems

实时和嵌入式系统在资源限制下工作，以执行计算任务并满足服务质量要求。而时间需求，以任务截止日期的形式，一直是实时系统建模和分析的核心，这类系统的系统级优化空间是丰富的。服务质量的某些度量（例如，解决方案的准确性）可能会被交换以满足其他约束。我们建议解决的两个制约因素是能源消耗和可靠性。从历史上看，这两个因素对计算系统都很重要，而且它们的重要性正在增长。能源消耗之所以重要，有两个原因：(i)许多嵌入式计算系统是电池供电的，需要有效地利用存储的能量；（ii）即使系统不是电池供电的，由于各种原因（包括传统能源的潜在短缺），对能源效率的关注也很重要。对可靠性的担忧是由CMOS晶体管尺寸的缩小引起的。这种技术进步增加了可以蚀刻在一个硅芯片上的计算核心的数量，从而提高了计算能力。计算能力的提高对实时和嵌入式应用程序以及个人计算机和数据中心都很有吸引力。另一方面，较小的晶体管更容易出现故障。我们相信软件开发人员不仅应该能够指定程序的正确版本，而且应该能够指定程序的近似版本。近似版本将节省能源并缩短执行时间。折衷的办法是根据近似版本计算出的数据的准确性。这样的近似可以在精度、能源效率和可靠性之间进行权衡。为了减少能源消耗或支持在时间限制内从故障中恢复，可以执行程序的近似版本。我们提出了一个研究计划来研究这种近似的价值，并建立一个框架，使系统能够通过牺牲可容忍范围内的正确性来适应能量和可靠性约束。

Real-time and embedded systems work under resource constraints to perform computational tasks and satisfy quality of service requirements. Whereas temporal requirements, in the form of task deadlines, have been central to the modelling and analysis of real-time systems, the space of system-level optimizations for such systems is rich. It is possible to trade some measures of service quality (e.g., accuracy of a solution) to satisfy other constraints. Two constraints that we propose to tackle are energy consumption and reliability. Both these factors have, historically, been important for computing systems and their significance is growing. Energy consumption is important for two reasons: (i) many embedded computing systems are battery-powered and need to utilize stored energy efficiently, and (ii) even when systems are not battery-operated, concerns about energy efficiency are important for various reasons including a potential shortfall in traditional energy sources. Reliability concerns are motivated by the shrinking size of CMOS transistors. This technological advance has increased the number of computational cores that can be etched onto one silicon die thereby increasing computational power. The increase in computational power is attractive for real-time and embedded applications as well as personal computers and data centres. On the flip side, smaller transistors are more vulnerable to faults. We believe that software developers should be able to specify not just a correct version of a program but approximate versions as well. Approximate versions would yield energy savings and shorter execution times. The compromise would be on the accuracy of the data computed by the approximate versions. Such approximations would then enable a tradeoff between accuracy, energy efficiency and reliability. To reduce energy consumption or to support recovery from faults within timing constraints, an approximate version of the program could be executed. We propose a research program for studying the value of such approximations and establishing a framework that would enable systems to adapt to energy and reliability constraints by sacrificing correctness within tolerable limits.