Abstract: Aiming at the problem of low efficiency and failure of dependent task offloading  in the cloud-edge-end architecture, we proposed a two-stage  dependent task offloading algorithm based on preference and virtual fitness. In the first stage, based on the proposed two-dimensional offloading preference factor,  direct offloading decisions were made for some sub-tasks of the dependent tasks, thus effectively reducing the size of the initial population of the genetic algorithm. In the second stage, we proposed a heuristic crossover method based on virtual fitness  to improve the crossover operator of  the fast non-dominated sorting genetic algorithm Ⅲ(NSGA-Ⅲ) based on reference points, which preserved the diversity of population and improved the convergence speed of the algorithm. Finally, we used  the improved algorithm to search for the optimal offloading decision set  for the subtasks of all dependent tasks. The experimental results show that compared with other algorithms, the proposed algorithm 
optimizes task completion time, task energy consumption and edge cloud cluster cost by 10.2%—18.3% on average and reduces the task failure rate by 10.7%—25.6% on average.

Key words: cloud-edge-end architecture, dependent task offloading, multi-objective optimization, virtual fitness, genetic algorithm