A Distributed Algorithm for Multi-Stage Computation Offloading

A scenario consisting of several mobile users, an access point (AP) and a cloud server in a multi-stage hierarchy is considered. Each user has a computation task which can be computed locally or offloaded to the AP or to the cloud server. Considering the shared access channel between users and the AP, the shared computation resources at the AP and the shared backhaul link connection from the AP to the cloud, an energy minimization computation offloading problem with a time constraint is tackled. The time constraint guarantees that the offloading time will not exceed the local computation time. In this paper, we propose a distributed game theoretic algorithm which decomposes the offloading problem into the subproblems of resource allocation and offloading decisions. The algorithm works iteratively between the two subproblems as follows: The AP receives offloading decisions from the users and accordingly optimizes the fractions of the bandwidth on the access channel, the fractions of the backhaul link rate and the fractions of the computation resource at the AP for all offloading users. Based on the assigned resources, each user autonomously decides between local computation or offloading to the AP or to the cloud server and reports its decision to the AP. Our proposed algorithm is shown to require only limited signaling between users and AP and converges in significantly few iterations. Furthermore, the results show that our algorithm performs close to the optimal policy.