Abstract: Workers in distributed machine learning often need to deal with heterogeneous tasks during the training process. However, the task publisher may not be able to determine which workers in the cluster of edge server (ES) are currently in training based on effective prior knowledge. To tackle the problem that the ES cluster cannot fulfill the maximization of the training performance and the quality of service at the same time, a scheduling algorithm of heterogeneous tasks is proposed. Firstly, the factors influencing the convergence performance of distributed training are analyzed under the constraints about cluster’s resources. Secondly, the optimization objective for maximizing training performance is established. Finally, the optimization problem is transformed into a multidimensional multiple-choice knapsack problem. The simulation results show that the proposed scheduling algorithm of heterogeneous tasks can maximize the performance of distributed training and simultaneously ensure the quality of ser- vice. 

Key words: distributed training, training performance, scheduling of heterogeneous tasks, multi- dimensional multiple-choice knapsack problem, convergence analysis