Abstract: With improving satellite computing capabilities, low earth orbit (LEO) satellites can be a supplement of ground networks. However, resource-limited LEO satellites face challenges in providing computationally intensive services as demand for computing power increases. Rapid movement of LEO satellites also poses a challenge in maintaining computation in the required area. To address these problems, this paper proposes a distributed computing strategy based on dynamic task decomposition. When making decisions of task decomposition, the strategy considers the resource utilization of the satellite network and the relationship between sub-tasks, and flexibly decomposes and aggregates tasks. Additionally, it proposes a long-term computing guarantee mechanism to keep the computation within user regions. The mechanism makes decisions based on real-time network topology and task decomposition relationship. It modularizes satellite handover process, and adjusts compression rate and service switching methods based on network conditions to minimize interruption during handover. The experimental results indicate that the strategy ensures long-term distributed computing. The average service time has increased by 110%, user satisfaction has improved about 20%, and both handover cost and inter-task communication cost have decreased by 15%.

Key words: satellite handover, computation residency, distributed computing, task decomposition

摘要： 低轨卫星具有覆盖范围广、离地面近等优势，随着在轨处理能力的不断增强，未来将成为地面网络的重要补充。然而，随着用户对网络服务实时性的需求日益增长，如何在资源有限的条件下，基于低轨卫星为用户提供计算密集型服务，已成为一个急需解决的问题。尤其是在低轨卫星高速移动、星间链路动态切换的情况下，如何保证空间计算能力能持续稳定地驻留在用户区域并提供稳定可靠的服务，无疑是一项巨大的挑战。为了解决上述问题，提出一种动态任务分解聚合的分布式计算策略，通过卫星分布式计算解决单星算力不足的问题。在进行任务分解与调度时，充分考虑卫星网络的资源占用情况以及子任务之间的关联关系，对任务进行灵活的分解聚合。此外，为将低轨卫星算力驻留在用户区域，解决低轨卫星服务周期短的问题，研究并设计了一种长时保障机制。根据实时卫星网络拓扑及任务分解调度图，结合任务间的关联关系进行迁移决策，对卫星迁移过程进行模块化设计，根据实时网络状况调整迁移过程中的数据压缩率以及服务切换方式，降低迁移过程中服务的中断时间。仿真实验结果表明，提出的策略可保障长时分布式计算，能提供服务的平均时长延长了110%，用户满意度提高了约20%，迁移开销以及任务间的传输开销均降低了约15%。

关键词: 卫星迁移, 算力驻留, 分布式计算, 任务分解