Adaptive Approximation-Aware Scheduling for Heterogeneous Computing Using Reinforcement Learning

Heterogeneous platforms are now key part of high-performance computing, offering improved performance and energy efficiency through the use of diverse compute resources. However, heterogeneity alone is often insufficient to meet the increasing demands for performance and energy savings. Approximation provides a complementary strategy, allowing controlled reductions in computational accuracy to further reduce execution time and energy consumption. Integrating approximation into heterogeneous systems, however, introduces additional complexity. It requires careful coordination of device selection, approximation levels, and code regions, all while managing trade-offs between performance, energy efficiency, and output quality. In this work, we propose a reinforcement learning (RL)-based scheduling framework that addresses this complexity by dynamically selecting both the execution device and the appropriate approximation strategy for each task. The scheduler supports optimization using traditional metrics like Energy-Delay Product (EDP), as well as a newly introduced unified metric that combines performance, energy efficiency, and accuracy loss, referred to as Energy-Delay-Loss Product (EDLP). Our evaluations show that RL-based approach outperforms state-of-the-art heuristics, enabling more adaptive and efficient execution of workloads on heterogeneous systems.