Abstract
The rapid advancements in heterogeneous parallel architectures, consisting of CPUs, GPUs, FPGAs, have introduced significant challenges in efficient resource management for high-performance computing systems. Static and heuristic-based approaches fail to address the adaptability required for handling varying workloads and hardware configurations which results in suboptimal performance and energy inefficiency. This research proposes a machine learning-driven adaptive resource management framework that dynamically optimizes task scheduling, resource allocation, and data placement. The framework employs regression models & reinforcement learning algorithms to predict workload behaviors, resource utilization, and task execution times in real time. Experimental results on heterogeneous testbed demonstrate a 21% reduction in task execution time, 18% improvement in energy efficiency, and 38% decrease in fault recovery time compared to conventional methods. These findings highlight the framework’s ability to improve resource utilization while maintaining reliability and minimizing energy overhead. The work advances the field by introducing a unified approach that integrates machine learning for runtime optimization across heterogeneous systems. Practical implications include its applicability to large-scale scientific simulations and deep learning tasks, where adaptive resource management is critical. Future study can focus on enhancing prediction accuracy by advanced deep learning techniques and extending the framework to handle emerging hardware accelerators and edge computing environments.