DuetEdge: Orchestrating service delay and resource consumption for multi-ESP task offloading

Edge computing systems increasingly support heterogeneous and delay-sensitive applications, such as augmented reality and autonomous driving. These applications generate computation tasks with diverse data volumes, computational intensities, and latency requirements. However, practical serving on resource-limited edge devices requires an explicit balance between stringent latency targets and infrastructure constraints. Consequently, managing the task streams makes online task offloading a critical systems problem. In dynamic multi-edge service provider (ESP) environments, we study multi-ESP task offloading through a service–resource orchestration perspective, since effective offloading must jointly account for short-term task service delay and long-term infrastructure resource consumption. To this end, we develop DuetEdge, a closed-loop service–resource orchestration framework for multi-ESP task offloading. DuetEdge integrates task-side heterogeneity, ESP-side resource dynamics, and online decision control to coordinate offloading decisions under evolving system conditions. We introduce an integrated utility score that unifies service delay and resource consumption, formulate the resulting service–resource orchestration problem, and prove that it is NP-hard. To solve this problem online, we devise TempoRL, a temporally aware deep reinforcement learning method for service–resource orchestration, with utility-aligned prioritized replay and an attention–long short-term memory (LSTM) dueling Q-network. TempoRL captures the long-term effects of sequential orchestration decisions on future workload evolution and infrastructure states. Our extensive evaluations under diverse edge scenarios demonstrate that DuetEdge significantly outperforms baselines, with average service delay and normalized resource consumption reduced by approximately 39% and 83%, respectively, in large-scale edge environments.