KIT | KIT-Bibliothek | Impressum | Datenschutz

Benchmarking Deep Learning Infrastructures by Means of TensorFlow and Containers

Grupp, Adrian; Kozlov, Valentin; Campos, Isabel; David, Mario; Gomes, Jorge; López García, Álvaro

Abstract (englisch):
Ever growing interest and usage of deep learning rises a question on the performance of various infrastructures suitable for training of neural networks. We present here our approach and first results of tests performed with TensorFlow Benchmarks which use best practices for multi-GPU and distributed training. We pack the Benchmarks in Docker containers and execute them by means of uDocker and Singularity container tools on a single machine and in the HPC environment. The Benchmarks comprise a number of convolutional neural network models run across synthetic data and e.g. the ImageNet dataset. For the same Nvidia K80 GPU card we achieve the same performance in terms of processed images per second and similar scalability between 1-2-4 GPUs as presented by the TensorFlow developers. We therefore do not obtain statistically significant overhead due to the usage of containers in the multi-GPU case, and the approach of using TF Benchmarks in a Docker container can be applied across various systems.

Open Access Logo


Download
Originalveröffentlichung
DOI: 10.1007/978-3-030-34356-9_36
Scopus
Zitationen: 1
Dimensions
Zitationen: 1
Zugehörige Institution(en) am KIT Steinbuch Centre for Computing (SCC)
Publikationstyp Buchaufsatz
Publikationsjahr 2019
Sprache Englisch
Identifikator ISBN: 978-3-030-34356-9
KITopen-ID: 1000100538
HGF-Programm 46.12.02 (POF III, LK 01) Data Activities
Erschienen in High Performance Computing : ISC High Performance 2019 International Workshops, Frankfurt, Germany, June 16-20, 2019. Ed.: M. Weiland
Auflage 1st ed.
Verlag Springer International Publishing
Seiten 478–489
Serie Theoretical Computer Science and General Issues ; 11887
Projektinformation DEEP-HybridDataCloud (EU, H2020, 777435)
Vorab online veröffentlicht am 03.12.2019
Schlagwörter Benchmarks, TensorFlow, ConvNet, Containers
Nachgewiesen in Scopus
Dimensions
KIT – Die Forschungsuniversität in der Helmholtz-Gemeinschaft
KITopen Landing Page