Deep-Dual-Learning-Based Cotask Processing in Multiaccess Edge Computing Systems

Abstract

Multiaccess edge computing (MEC) systems provide low-latency computing services for Internet of Things (IoT) applications by processing IoT data on edge servers. In the era of heterogeneous IoT environments, the success of IoT applications hinges on the processing of diversified IoT data. To leverage MEC systems to enable timely IoT services, we characterize IoT applications as cotasks, where each cotask is completed only if all its constituent subtasks (e.g., IoT data processing) are finished. Existing works have been devoted to the design of task offloading and scheduling decisions for MEC-enabled IoT applications, but they mostly neglect the cotask feature. In this article, we investigate the problem of cotask processing in MEC systems, and we formulate it as a nonlinear program (NLP) to minimize total cotask completion time (TCCT). In the light of uncertain communication latency, we transform the NLP to a parameterized and unconstrained version, based on which we propose the deep dual learning (DDL) method, where the learner keeps updating primal and dual variables based on randomly perturbed samples. Furthermore, we provide the duality gap and time complexity analyses for the DDL method. Our simulation results demonstrate that the proposed solution can gradually converge over iterations, and its TCCT performance outperforms other comparison schemes under various system settings.