Dynamic scheduling can improve the latency and memory efficiency of deep neural network inference on edge devices, but it often introduces cold-start overhead when a newly deployed model requires online profiling and policy adaptation before reaching stable performance. This paper proposes EdgeOpt-Sched-CS, a cold-start-aware extension of dynamic graph scheduling for edge inference. The key idea is to initialize the scheduler of a target computation graph using scheduling knowledge transferred from structurally similar source graphs, instead of starting from a generic policy. EdgeOpt-Sched-CS constructs compact graph signatures, retrieves relevant source schedulers, and performs lightweight cold-start-aware online adaptation during early deployment. We evaluate the framework across representative device–model scenarios involving lightweight convolutional neural networks, transformer models, and quantized language-model workloads. The results show that EdgeOpt-Sched-CS reduces cumulative cold-start latency by 10.6–20.4% and shortens time-to-stability by 5.2–21.7%, while preserving the steady-state latency–memory behavior of the original dynamic scheduler with only small additional scheduling overhead. These findings indicate that scheduler initialization is an important optimization dimension for adaptive edge inference and that prior scheduling knowledge can be effectively reused across related computation graphs.
谷玉昌 et al. (Fri,) studied this question.