Task Offloading Strategy for LEO Satellite Networks Based on Complex Dependency Decomposition

地面设备受限于本地计算资源，需将计算任务卸载至低轨卫星网络以降低处理时延。在任务卸载过程中，为了突破单个低轨卫星的计算资源限制，则需要将计算任务拆解成多个子任务。然而任务内部复杂的依赖关系给任务分解带来极大挑战。为此，文章提出了一种基于复杂依赖关系分解的任务卸载策略。首先，利用区域控制器将网络划分为控制域，统筹协调域内外计算资源，并负责任务卸载。其次，将具有复杂依赖关系的任务建模为有向无环图(DAG)，并采用图卷积网络(GCN)对复杂依赖关系进行特征提取与分解。然后，将任务卸载问题构建为马尔可夫决策过程(MDP)，利用深度Q网络(DQN)算法求解最优卸载策略。最后，实验结果表明，该策略能够有效降低任务处理时延。Due to the limitations of local computing resources on ground devices, computation tasks are offloaded to Low Earth Orbit (LEO) satellite networks to reduce processing delays. During the task offloading process, to overcome the computing resource constraints of single LEO satellites, a complex task needs to be decomposed into multiple subtasks. However, complex dependency relationships within the tasks pose significant challenges for task decomposition. To address this problem, this paper proposes a task-offloading strategy based on the decomposition of complex dependency relationships. Firstly, a regional controller is introduced to divide a network into control domains, which coordinates computing resources within and across domains in a unified manner and is responsible for task offloading. Secondly, tasks with complex dependencies are modeled as directed acyclic graphs (DAG), and graph convolutional networks (GCN) are utilized to model and extract features from these complex dependencies. Then, the task offloading problem is formulated as a Markov decision process (MDP), and the deep Q network (DQN) algorithm is used to search for the optimal offloading strategy. Finally, experimental results demonstrate that this strategy can effectively reduce the delay in task processing.