Training DNN Models over Heterogeneous Clusters with Optimal Performance