CPU-GPU Heterogeneous Computation Offloading and Resource Allocation Scheme for Industrial Internet of Things

Zixuan He,Yanjing Sun,Bowen Wang,Song Li,Beibei Zhang

DOI: https://doi.org/10.1109/jiot.2023.3332748

IF: 10.6

2023-01-01

IEEE Internet of Things Journal

Abstract:The computing process of tasks in Industrial Internet of Things (IIoT) environments is becoming increasingly complex due to the development of 5G and artificial intelligence. Leading devices are increasingly relying on heterogeneous platforms that integrate different types of processing units, such as CPUs, GPUs, and other resources, to meet the requirements of delay-sensitive and computing-intensive tasks. However, compared to conventional general-proposed CPU computing, CPU-GPU heterogeneous computing typically involves three processes, i.e., task preprocessing, hybrid computing, and result aggregation. These processes are associated with particular computing resources, which increases the difficulty of task offloading and computing resource allocation under task-specific resource and delay constraints. In this article, we first propose a three-stage heterogeneous computing (TSHC) model to practically describe the computing process of parallelizable tasks. Considering the heterogeneous computing resources, device queue backlogs, and collaboration of multiple edge servers, the joint task offloading and heterogeneous resource allocation (JCOHRA) problem is formulated to minimize the long-term average delay of tasks. Then, the Lyapunov optimization method is adopted to simplify the long-term queue stability constraint to a single-slot dynamic optimization problem, which is then modeled as a Markov decision process (MDP). Owing to the tight coupling between decision variables and enormous action space, we propose the multihead proximal policy optimization (MH-PPO)-based JCOHRA algorithm, which is enabled by elaborate constraint transformation and reward function design. Simulation results demonstrate that the JCOHRA scheme achieves better performance than baseline methods in minimizing the long-term average delay of tasks.