Hard real-time application mapping reconfiguration for NoC-based many-core systems

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.