ECO-EDGE: joint DAG offloading and fabric routing in Multi-AP MEC via conic optimization

Abstract

Mobile Edge Computing (MEC) enables low-latency services by migrating computational resources closer to end-users. However, as mobile workloads increasingly adopt complex Directed Acyclic Graph (DAG) topologies, system performance is strongly affected by interdependent data shuffling across the inter-server fabric and synchronization constraints during uplink transmission. Existing models frequently neglect physical fabric capacity limits and decouple communication from computation, leading to inefficient energy usage and increased latency. In this paper, we propose a coupled Mixed-Integer Non-Linear Programming (MINLP) formulation that jointly optimizes multi-AP association, task placement, Dynamic Voltage and Frequency Scaling (DVFS), and explicit inter-server routing under per-application deadline constraints. To address the inherent non-convexity and combinatorial complexity of the joint problem, we introduce the ECO-EDGE framework. ECO-EDGE decomposes the original MINLP using a Block Coordinate Descent (BCD) architecture and solves relaxed subproblems via tailored convexification techniques, including McCormick-based convex envelopes for bilinear placement products, a “Bits-per-RB” formulation that maps communication constraints into exponential cones, and power- and rotated second-order cone representations for computation and fabric delay modeling. The resulting relaxed BCD scheme exhibits monotonic objective decrease under exact block solution, while deterministic rounding and topological list-scheduling are applied to recover feasible schedules, which are subsequently verified for deadline satisfaction. Simulation results in dense multi-AP deployments show that ECO-EDGE reduces total system energy by up to \(81.8\%\) compared to a Local-Only baseline and achieves an empirical deadline satisfaction rate of \(97.9\%\) , outperforming heuristic, fabric-oblivious, and disjoint optimization baselines.