APT-SAT: An Adaptive DNN Partitioning and Task Offloading Framework Within Collaborative Satellite Computing Environments

Shifeng Peng, Zhishu Shen, Qiushi Zheng, Xuefeng Hou, Dawen Jiang, Jingling Yuan, Jiong Jin

Abstract: Satellite computing has emerged as a promising technology for next-generation wireless networks, providing various data processing capabilities within space networks. This advancement facilitates the widespread implementation of artificial intelligence.(AI)-driven IoT applications like image processing tasks involving deep neural networks (DNN). However, the constrained computational and communication capacities of individual satellites pose significant challenges in efficiently managing DNN tasks generated by diverse users. One viable solution involves partitioning DNN tasks into multiple subtasks and distributing them across multiple satellites for collaborative computing. Despite its potential, it faces challenges in effectively partitioning DNNs to minimize delay and energy consumption while ensuring efficient subtask allocation to maintain load balancing and high task completion rates. To this end, we propose an adaptive DNN partitioning and task offloading framework named APT-SAT to enable efficient distributed computing among multiple satellites. APT-SAT incorporates an adaptive DNN partitioning algorithm that aims to evenly distribute the workload of DNN slices. Furthermore, a routing and task offloading algorithm, leveraging the soft actor-critic (SAC) approach, is introduced to optimize offloading decisions. The extensive experiments validate the effectiveness of APT-SAT in terms of task completion rate, delay, energy consumption, and resource utilization, highlighting its potential for enabling efficient satellite-based AI applications.

External IDs:dblp:journals/tnse/PengSZHJYJ26