Joint Request Offloading and Resource Allocation for Long-Term Utility Optimization in Collaborative Edge Inference With Time-Coupled Resources

Extensive research on edge inference has devoted in optimizing service performance for users. However, recent studies have overlooked the desired utility of application service provider (ASP), which is crucial for achieving long-term service provisioning. Besides, efficient request offloading and resource allocation are essential for optimizing long-term utility of ASP in dynamic networks with time-coupled resources. To address these issues, this paper formulates a long-term utility optimization problem in collaborative edge inference system. The objective is to maximize the long-term average utility of ASP, by jointly optimizing request offloading and resource allocation, under the time-coupled resource constraints. To solve the problem, a Lyapunov based online algorithm is proposed to decompose it into a series of one-slot deterministic problems by decoupling the time-coupled resource constraints. Only the current network states are required for one-slot problem. Then, the one-slot problem is converted into a master request offloading problem with an inner resource allocation problem. A distributed algorithm is proposed to derive the optimal decision to inner problem, while a coalition based algorithm is proposed to seek the stable solution to master problem. Experimental results show that, the proposed algorithm outperforms baseline algorithms for most cases, in terms of long-term average utility of ASP.