Hierarchical Offloading Optimization for Collaborative DNN Inference in Satellite Edge Computing Networks

Satellite Edge Computing (SEC) enables intelligent onboard processing, which is essential for mission-critical tasks involving high data capture rates and compute-intensive Deep Neural Network (DNN) models. However, limited satellite computing capacity often causes significant processing delays, underscoring the need for efficient collaboration across the SEC network to ensure real-time responsiveness and optimal resource utilization. Since DNN models can be partitioned into sub-tasks executed in parallel across multiple nodes, partitioning naturally facilitates collaborative processing by exploiting idle resources and reducing transmission overhead, while also accelerating inference. Motivated by this, we propose HiO2, a hierarchical task offloading framework that maximizes system throughput by coordinating partitioned sub-tasks across satellites and ground stations. To realize this hierarchical design, HiO2 incorporates two key mechanisms: (i) a swarm-level offloading strategy that assigns tasks to swarms according to their optimal allocation, and (ii) a node-level runtime-aware partitioning and sub-task offloading scheme that dynamically determines both the number of partitions and the placements of cut-points, while coordinating the execution of sub-tasks across nodes based on workload and network conditions. Extensive experiments on a prototype system demonstrate that HiO2 not only outperforms state-of-the-art methods in system throughput and average task response time but also consistently meets real-time task deadlines.