Real-Time Task Migration for Dynamic Resource Management in Many-Core Systems

Abstract

Dynamic resource management strategies in embedded many-core systems rely on task migration to adapt the deployment (mapping) of applications dynamically, e.g., for thermal/power management or load balancing. In case of hard real-time applications, however, the current practice of on-line application adaptation is limited to reconfiguring the whole application between a set of statically computed mappings with statically verified timing guarantees. This heavily restricts the application’s adaptability. To enable hard real-time task migrations in many-core systems without relying on a static analysis, this paper presents (i) a predictable task migration mechanism supported with (ii) a lightweight migration timing analysis and (iii) a lightweight migration timing feasibility check which can be applied on-line to bound on the worst-case temporal overhead of a migration and examine the admissibility of this overhead w.r.t. the hard real-time requirements of the application. For a variety of applications and many-core platforms, we experimentally demonstrate the feasibility of hard real-time task migrations, the lightness of the proposed timing analysis and feasibility check for on-line use, and the advantage of the proposed task migration approach over mapping reconfiguration as the state-of-the-art real-time adaptation approach for many-core systems.