An Analysis of GPU Sharing Methods for Deep Learning Applications in a Heterogeneous Cluster Environment