Task-Oriented Based Computing Power Measurement for Task Scheduling in Computing Power Networks

Computing Power Network (CPN) has emerged as a promising paradigm that integrates communication networks with distributed computing resources to support new-generation applications, such as Artificial Intelligence (AI) and Virtual Reality (VR). While CPN enables efficient utilization of heterogeneous computing resources, its implementation faces several challenges, including accurate computing power measurement (CPM) of heterogeneous computing resources and efficient task scheduling. To adapt computing tasks with different attributes to heterogeneous computing resources, we propose a novel framework for CPM and computing task scheduling. First, to facilitate accurate CPM, a clustering-based algorithm categorizing tasks into distinct types based on their attributes is proposed. Then, a neural network model is designed to measure the computing power of heterogeneous computing resources after the task clustering. Second, we formulate a computing task scheduling problem to maximize the matching degree between tasks and computing resources based on the proposed CPM, whose environment is modeled as a Markov Decision Process (MDP). Then, a deep reinforcement learning-based task scheduling strategy with an adaptive exploration mechanism to accelerate the convergence is proposed. Simulation results show that the proposed framework can improve the rewards of task scheduling by up to $30.4 \%$ compared to baselines.