Topology-aware job scheduling for machine learning cluster

Parameter Server (PS) has been widely used to train a large amount of data on multiple machines in parallel. In parameter server, a critical problem is how to effectively schedule multiple training jobs to minimize the job completion time. Some existing work has proposed methods of setting the number of concurrent workers. However, they do not effectively consider the topology of GPU placement which affects the efficiency of communication. This paper proposes a novel resource-to-time model based on the number of workers and the topology of GPU placement. According to the model, we propose an algorithm called TOPO-PS particularly for topology problem in parameter servers. The algorithm achieves the placement strategy based on graph mapping algorithm. Evaluation under various algorithms evidences the superiority of our algorithm. TOPO-PS yields shorter job completion, by up to 53.48% of that of FIFO and 88.77% of OASIS.