Prediction-Assisted Online Distributed Deep Learning Workload Scheduling in GPU Clusters