Multi-Agent DRL-Based Large-Scale Heterogeneous Task Offloading for Dynamic IoT Systems

In dynamic IoT system, the device may generate multiple heterogeneous computational tasks, that require CPU and GPU co-processing, in each period. Furthermore, different heterogeneous computing tasks have specific requirements for GPU resource types. Realizing real-time scheduling and processing of large-scale hybrid computing tasks with high heterogeneity and dense quantity has become an urgent problem. First, we propose a cloud-based task processing framework that uses multi-level feedback queues to ensure the fairness of large-scale task parallel computing. Second, we decoupled the original problem into a series of mixed-integer nonlinear programming problems using Lyapunov optimization, aiming to reduce the solution complexity of the real-time scheduling problem. Finally, we propose a multi-agent reinforcement learning algorithm, employing long and short-term memory networks with parameter resetting, to generate task offloading decisions in near real-time based on partially knowable future information. Through extensive simulation experiments, we have demonstrated that our algorithm can reduce the average task processing time by approximately 19.95% and enhance the task processing capability of the IoT system by roughly 12.43%, especially in large-scale hybrid computing task systems.