EdgeOpt-Sched: A Dynamic GNN and RL Scheduler for DNN Acceleration on Edge Devices

The growing adoption of edge AI in applications such as smart manufacturing and autonomous systems demands efficient deep neural network (DNN) inference on resource-constrained devices. However, mainstream intermediate representations (IRs), including ONNX and TVM Relay, rely heavily on static scheduling, which limits adaptability to the dynamic and heterogeneous nature of edge environments. To address this issue, we propose EdgeOpt-Sched, a dynamic operator scheduling framework that leverages Graph Neural Networks (GNNs) and Reinforcement Learning (RL) to optimize model inference graphs. Specifically, we introduce TopoGNN for modeling operator dependencies and PPO-Sched to learn adaptive scheduling strategies via Proximal Policy Optimization. Experimental results on edge devices using BERT, ResNet50 and Vision Transformer demonstrate that EdgeOpt-Sched reduces average inference latency by 31.4% and improves memory utilization by 26.2% over existing IR-based runtimes. Our framework offers a portable and lightweight solution for efficient DNN deployment across diverse edge platforms. We compare against state-of-the-art schedulers including TVM MetaSchedule and IOS and observe consistent end-to-end latency and throughput gains across BERT, ViT and ResNet workloads on ARMv8 edge devices. Compared with TVM MetaSchedule (offline search) and IOS (operator-group DP), our method reduces latency while lowering peak memory and remains stable under runtime variability (CPU load, thermal throttling).