Design Methodology and Run-Time Management for Predictable Many-Core Systems

摘要

Many-core systems provide a feasible means to build high-performance multi-application systems. They are increasingly exposed to dynamic changes due to varying and online modified application mixes, as well as unavailability of hardware resources due to thermal and power management or faults. Particularly, when applications with real-time requirements are executed, these constraints may become a problem. Here, self-adaptive mechanisms are a means to optimally and feasibly manage such a system at run-time. This paper gives an overview of a design flow for hybrid application mapping for predictable many-core systems. The idea behind this concept is to perform a complex timing analysis and verification of an application during design time and then exploit this information for RM of the system. We propose a novel optimization algorithm to perform RM, which statically calculates a set of application mappings that are guaranteed to fulfill real-time requirements while heuristically optimizing the system objectives. Here, minimization of the energy consumption is exemplary chosen. The algorithm achieves significantly better results than a state-of-the-art approach, especially for systems with a high resource congestion.