Towards A Resource Efficient Framework For Distributed Deep Learning Applications