Communication-Dependent Computing Resource Management for Concurrent Task Orchestration in IoT Systems

Abstract

Recent advances in distributed machine learning and wireless network technologies are bringing new opportunities for Internet of Things (IoT) systems, where smart devices are often wirelessly connected to collaborate, jointly completing tasks known as communication-dependent computing (CDC) tasks. However, due to the dependence of computing on communication and the presence of concurrent tasks, it remains a challenge to optimize CDC task performance and efficiency while fulfilling multi-dimensional requirements, particularly with incomplete system information and dynamic environmental impacts. To overcome these, we present a concurrent CDC task framework to model the correlated communication and computing stages and multi-dimensional requirements of CDC tasks. We then formulate a task orchestration and resource management problem to optimize overall utility, where each task's utility is designed as a joint metric including the cumulative computing deviation and time efficiency of task completion. To solve this, we employ auxiliary graphs to capture the topological information of tasks and resources, and update weights based on the utility in dynamic environments. Subsequently, a multi-agent reinforcement learning algorithm is leveraged to make distributed decisions with incomplete information. Experiments demonstrate the proposed approach outperforms baselines in terms of task performance and efficiency, indicating our solution holds great potential.