Task Offloading with Dual-Mode Switching in Multi-access Edge Computing

Task offloading is an important mechanism in edge computing that can reduce the execution latency and dropping rate of computation tasks. However, the existing deep learning-based task offloading methods generate offloading decisions with high randomness and poor quality in the early stage before the model converges. The bid-based task offloading methods, in turn, have difficulty in utilizing task execution history information to guide offloading decisions. To overcome the shortcomings of the existing task offloading methods, we design a dual-mode switching method for task offloading in the multi-access edge computing environments. The method dynamically switches between the deep reinforcement learning-based decision mode and the dynamic bidding-based decision mode. The method utilizes a global bidding mechanism to fine tune the raw offloading decisions made by the two modes to reduce doom-to-fail task offloading decisions. The experimental results on the simulator show that the proposed dual-mode switching task offloading method is able to achieve a low task dropping rate across the entire execution process. The average execution latency of the tasks gradually decreases and converges as time grows.