ARC Containers for AI Workloads: Singularity Performance Overhead
Document Type
Conference Proceeding
Publication Date
7-28-2019
Abstract
Containerization has taken the software world by storm. Deployment complications, like requiring elevated (i.e. "root") permissions to run, have slowed the adoption of containers in shared advanced research computing (ARC) environments. Singularity is a containerization approach that is designed for ARC in shared high performance computing (HPC) clusters. With the creation of the Singularity, there is finally a viable scientific container solution. However very few papers have looked at the performance trade-offs of deploying applications using a container based model. The authors are not aware of any published studies evaluating the trade-offs of the deployment models with complex Artificial Intelligence (AI) workloads. Without detailed evaluations of the performance trade-offs scientists and engineers are unable to make an informed decision on deployment model for time sensitive training or low power inference. Furthering previous research in this area and using emerging community developed benchmarks, we examine performance trade-offs of running AI workloads in a containerized Singularity environment.
Source Publication
Proceedings of the Practice and Experience in Advanced Research Computing on Rise of the Machines (learning) (ISBN 978-1-4503-7227-5)
Recommended Citation
Marvin Newlin, Kyle Smathers, and Mark E. DeYoung. 2019. ARC Containers for AI Workloads: Singularity Performance Overhead. In Practice and Experience in Advanced Research Computing 2019: Rise of the Machines (learning) (PEARC '19). Association for Computing Machinery, New York, NY, USA, Article 1, 1–8. https://doi.org/10.1145/3332186.3333048
Comments
Copyright © 2019 Public Domain.
This paper is authored by employees of the United States Government and is in the public domain. Non-exclusive copying or redistribution is allowed, provided that the article citation is given and the authors and agency are clearly identified as its source.
The "Link to Full Text" on this page opens the full conference paper hosted at ACM.
Co-authors M. Newlin (AFIT-ENG-MS-20-M-048) and K. Smathers (AFIT-ENG-MS-20-M-062) were enrolled in an AFIT graduate program at the time of this conference. (March 2020 graduates)