Real-time applications are increasingly targeting many-core platforms, demanding predictability in a highly dynamic environment. To enable this shift, for each application, a set of mapping candidates with diverse resource requirements and performance qualities (latency, energy, etc.) may be computed at design time, and subsequently, exploited at run time to launch the application on a mapping that adheres to the on-line quality and resource constraints. These constraints, however, may also change during execution such that the mapping in use fails to satisfy them, necessitating a switch to another mapping. This process, namely, mapping reconfiguration, involves the migration of several tasks and may harm timing predictability if the reconfiguration overhead is not accounted for. This paper presents a deterministic mapping reconfiguration methodology to enable predictable reconfigurations among a given set of mappings. To this end, first in an off-line analysis, we (a) identify low-latency migration routes with minimal allocation overhead for each pair of source/target mappings and (b) bound the worst-case reconfiguration latency using an off-line timing analysis. This information is then used at run time to perform timely reconfigurations. We further investigate a (c) hybrid timing analysis which regards the actual availability of communication resources at run time to derive tighter latency bounds. Experimental results for a variety of applications show that the proposed methodology enables reconfigurations with low allocation overhead and affordable latency. To demonstrates the practicality of the proposed methodology and the advantages of the hybrid latency analysis over its off-line counterpart, we present a case study on thermal management of many-core systems using mapping reconfiguration.

中文翻译：

---

基于 NoC 的众核系统的硬实时应用程序映射重新配置

实时应用程序越来越多地面向众核平台，要求在高度动态的环境中具有可预测性。为了实现这种转变，对于每个应用程序，可以在设计时计算一组具有不同资源需求和性能质量（延迟、能量等）的候选映射，然后在运行时利用这些映射在映射上启动应用程序坚持在线质量和资源限制。然而，这些约束也可能在执行过程中发生变化，以至于使用中的映射无法满足它们，从而需要切换到另一个映射。这个过程，即映射重配置，涉及多个任务的迁移，如果不考虑重配置开销，可能会损害时序可预测性。本文提出了一种确定性映射重新配置方法，以在给定的一组映射之间实现可预测的重新配置。为此，首先在离线分析中，我们 (a) 为每对源/目标映射确定具有最小分配开销的低延迟迁移路线和 (b) 使用离线限制最坏情况的重新配置延迟时序分析。然后在运行时使用此信息来执行及时的重新配置。我们进一步研究了 (c) 混合时序分析，该分析考虑了运行时通信资源的实际可用性，以获得更严格的延迟界限。各种应用的实验结果表明，所提出的方法能够以低分配开销和可承受的延迟进行重新配置。