Energy-efficient execution of concurrent applications in heterogeneous many-core systems

近年来，电源效率已经成为嵌入式系统研究的前沿，并且被认为与性能相比具有同等或更重要的意义。由于几乎没有迹象表明商业能源存储解决方案能够跟上硅和计算机架构创新的步伐，这一趋势可能会持续下去。文献中已经提出了多核和多核系统的运行时管理，以有效地应用动态电压和频率缩放（DVFS）和动态电源管理（DPM）来权衡更高的核心利用率以降低功耗，由于硬或软截止日期仍然坚持，对性能的影响很小或没有影响。这些方法倾向于在某些情况下只考虑单个应用程序的执行，或者在其他情况下考虑一系列顺序（即非并发）应用程序的执行。众所周知，应用程序并发性具有多个相关的挑战，通过DVFS扩展添加运行时电源管理进一步增加了对有限和受限资源的竞争，并减少了基于截止日期的实时系统的时间余量。如果计算任务在设计时已知，或者具有周期性，则可以确定简单的解决方案来解决这一挑战。然而，嵌入式系统通常处理非周期性任务（例如涉及用户输入和异步网络调用的任务），这意味着必须在运行时做出决策，以便以节能的方式满足最后期限。本研究将研究运行时DVFS管理和非周期任务调度之间的相互作用，以确定运行时管理的哪些元素可以在现实世界的多核异构系统中合理地实现。预期的贡献将是一种在运行并发非周期性应用程序时管理异构多核系统的功耗的方法。这种方法将通过数学建模和开发板（如Odroid-XU4和NVIDIA Jetson）上的实际实验进行全面评估。

) In recent years, power efficiency has come to the forefront of research in embedded systems, and has been considered to have arguably equal or more importance when compared to performance. As there are few indications that commercial energy storage solutions will keep pace with innovation in silicon and computer architecture, this trend is likely to continue. Run-time management of multi- and many-core systems have been proposed in the literature to effectively apply dynamic voltage and frequency scaling (DVFS) and dynamic power management (DPM) to trade-off higher core utilisation for reduced power consumption, with little or no impact on performance as hard or soft deadlines remain adhered to. These approaches have tended to consider the execution of, in some cases, just a single application or, in other cases, a series of sequential (i.e. non-concurrent) applications. Application concurrency is known to have multiple associated challenges, and adding run-time power management through DVFS scaling further increases competition for limited and constrained resources, and reduces the time margin in a deadline based real-time system. If the computation tasks are known at design time, or are of a periodic nature, simple solutions can be identified to solve this challenge. However, embedded systems typically deal with aperiodic tasks (such as those involving user input and asynchronous network calls) which mean that decisions must be made at run-time in order to meet deadlines in a power-efficient manner. This research will investigate the interaction between run-time DVFS management and aperiodic task scheduling to determine what elements of run-time management can plausibly be implemented in a real-world multi-core heterogeneous system. Anticipated contributions will be an approach to managing the power consumption of a heterogeneous many-core system, at run-time, while it is running concurrent aperiodic applications. This approach(es) will be thorough evaluated through mathematical modelling and practical experimentation on development boards such as the Odroid-XU4 and the NVIDIA Jetson.