Using Emulab for Deep Learning Performance Comparisons among Network Topologies

Abstract

Emulab is a versatile research framework proposed and implemented by Utah University, instantly providing a dedicated cluster system using real systems and switches upon a user request. As machine learning has been used in most areas, it is also natural to try to use Emulab for machine learning. However, there has been no study on how to configure Emulab nodes to run machine learning. In particular, our research focuses on comparing the performance of TensorFlow, one of the most widely used tools in deep learning field, to find out which network topology has the best performance. In our experiments, the star topology cluster running the data parallelism model of TensorFlow showed the best performance. In addition, to our best knowledge, this study is the first research to investigate to figure out the relationship between cluster interconnects and deep learning performance by using Emulab.