HADES: Heterogeneity-Aware Dependent Task Offloading in Collaborative Edge Networks

Abstract

Regarding the computational intensity and stringent latency requirements of modern applications comprised of many dependent subtasks, collaborative edge computing (CEC) emerges as a solution to guarantee the service-level objectives (SLOs) by offloading subtasks to multiple distributed edge servers. However, existing offloading schemes remain inadequate in addressing the stochastic task arrival and diverse task-resource affinity that stems from computational efficiency differences when executing various tasks on heterogeneous resources, leading to suboptimal system performance. Moreover, due to continuous updates in both applications and resources (e.g., hardware and runtime environment) in practical environments, comprehensively evaluating task performance profiles across all resource configurations becomes prohibitively expensive. To address these challenges, we propose heterogeneity-aware dependent task offloading scheme (HADES). We first formulate the problem of minimizing the long-term system energy consumption while guaranteeing average application latency. To optimize task offloading policy without relying on prior knowledge of task-resource affinity, HADES adopts a novel generative diffusion model-based graph reinforcement learning approach. Specifically, HADES employs graph neural networks to handle the graph-structured states of both the edge network and dependent applications, and leverages diffusion model-based policy representation for efficient policy exploration in a large action space. In addition, a Sortino ratio-based latency reward is designed for HADES to achieve risk-aware latency optimization. Extensive evaluations on real-world datasets demonstrate that HADES can reduce the latency and energy consumption by at most 98.1% and 86.9% compared to existing baselines.