Benchmarking Virtual Machine and Container-based Services for DNN Training
Journal
2021 29th Telecommunications Forum (TELFOR)
Date Issued
2021-11-23
Author(s)
Postolovski, Damjan
Kjirovski, Kiril
DOI
10.1109/telfor52709.2021.9653280
Abstract
Containerized infrastructure is becoming the leading strategy for deployment and operation of software components including trending neural networks. Training neural networks requires computational resources and complex software environments with a lot of dependencies. In this regard, containers can solve some of the software related challenges. We perform benchmarks to compare the performances of containers inside a virtual machine and a bare metal model for training three artificial neural networks on two leading cloud providers. The results show negligible performance drawback for using containers in the stack tested on Google Cloud and clear performance advantages in Amazon Web Services.
Subjects