Analysis and Comparison of Distributed Training Techniques for Deep Neural Networks in a Dynamic Environment
Ermias Gebremeskel · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2018
Deep learning models' prediction accuracy tends to improve with the size of the model. The implications being that the amount of computational power needed to train models is continuously increasing. Distributed deep learning training tries to address this issue by spreading the computational load onto several devices. In theory, distributing computation onto N devices should give a performance improvement of xN. Yet, in reality the performance improvement is rarely xN, due to communication and other overheads. This thesis will study the communication overhead incurred when distributing deep learning training. Hopsworks is a platform designed for data science. The purpose of this work is to explore a feasible way of deploying distributed deep learning training on a shared cluster and analyzing the performance of different distributed deep learning algorithms to be used on this platform. The findings of this study show that bandwidth-optimal communication algorithms like ring all-reduce scales better than many-to-one communication algorithms like parameter server, but were less fault tolerant. Furthermore, system usage statistics collected revealed a network bottleneck when training is distributed on multiple machines. This work also shows that it is possible to run MPI on a hadoop cluster by building a prototype that orchestrates resource allocation, deployment, and monitoring of MPI based training jobs. Even though the experiments did not cover different cluster configurations, the results are still relevant in showing what considerations need to be made when distributing deep learning training.