Thermal Load-aware Adaptive Scheduling for Heterogeneous Platforms

Modern-day heterogeneous embedded computing platforms integrate processing elements (PE) with varying compute capabilities on the same die. While such platforms help in delivering high-throughput computation with low power budgets, it also exposes the possibility of violating the thermal envelope. Sustained thermal envelope violation degrades the reliability of the PEs. This work presents an OpenCL run-time extension which adaptively tunes parameters of heterogeneous multicores in order to respect a given temperature constraint without violating real-time deadlines. The feedback-oriented iterative behavior of the proposed run-time extension helps in cancelling out core-level temperature constraint violations which may happen due to dynamic task injection in plug-n-play embedded computing platforms.