Abstract
Scaling CNN training is necessary to keep up with growing datasets and reduce training time. We also see an emerging need to handle datasets with very large samples, where memory requirements for training are large. Existing training frameworks use a data-parallel approach that partitions samples within a mini-batch, but limits to scaling the minibatch size and memory consumption makes this untenable for large samples. We describe and implement new approaches to convolution, which parallelize using spatial decomposition or a combination of sample and spatial decomposition. This introduces many performance knobs for a network, so we develop a performance model for CNNs and present a method for using it to automatically determine efficient parallelization strategies. We evaluate our algorithms with microbenchmarks and image classification with ResNet-50. Our algorithms allow us to prototype a model for a mesh-tangling dataset, where sample sizes are very large. We show that our parallelization achieves excellent strong and weak scaling and enables training for previously unreachable datasets.
| Original language | English (US) |
|---|---|
| Title of host publication | Proceedings - 2019 IEEE 33rd International Parallel and Distributed Processing Symposium, IPDPS 2019 |
| Publisher | Institute of Electrical and Electronics Engineers Inc. |
| Pages | 210-220 |
| Number of pages | 11 |
| ISBN (Electronic) | 9781728112466 |
| DOIs | |
| State | Published - May 2019 |
| Event | 33rd IEEE International Parallel and Distributed Processing Symposium, IPDPS 2019 - Rio de Janeiro, Brazil Duration: May 20 2019 → May 24 2019 |
Publication series
| Name | Proceedings - 2019 IEEE 33rd International Parallel and Distributed Processing Symposium, IPDPS 2019 |
|---|
Conference
| Conference | 33rd IEEE International Parallel and Distributed Processing Symposium, IPDPS 2019 |
|---|---|
| Country | Brazil |
| City | Rio de Janeiro |
| Period | 5/20/19 → 5/24/19 |
Fingerprint
Keywords
- Algorithms
- Convolution
- Deep learning
- HPC
- Performance modeling
ASJC Scopus subject areas
- Computer Networks and Communications
- Hardware and Architecture
- Information Systems and Management
Cite this
Improving strong-scaling of CNN training by exploiting finer-grained parallelism. / Dryden, Nikoli; Maruyama, Naoya; Benson, Tom; Moon, Tim; Snir, Marc; Van Essen, Brian.
Proceedings - 2019 IEEE 33rd International Parallel and Distributed Processing Symposium, IPDPS 2019. Institute of Electrical and Electronics Engineers Inc., 2019. p. 210-220 8820780 (Proceedings - 2019 IEEE 33rd International Parallel and Distributed Processing Symposium, IPDPS 2019).Research output: Chapter in Book/Report/Conference proceeding › Conference contribution
}
TY - GEN
T1 - Improving strong-scaling of CNN training by exploiting finer-grained parallelism
AU - Dryden, Nikoli
AU - Maruyama, Naoya
AU - Benson, Tom
AU - Moon, Tim
AU - Snir, Marc
AU - Van Essen, Brian
PY - 2019/5
Y1 - 2019/5
N2 - Scaling CNN training is necessary to keep up with growing datasets and reduce training time. We also see an emerging need to handle datasets with very large samples, where memory requirements for training are large. Existing training frameworks use a data-parallel approach that partitions samples within a mini-batch, but limits to scaling the minibatch size and memory consumption makes this untenable for large samples. We describe and implement new approaches to convolution, which parallelize using spatial decomposition or a combination of sample and spatial decomposition. This introduces many performance knobs for a network, so we develop a performance model for CNNs and present a method for using it to automatically determine efficient parallelization strategies. We evaluate our algorithms with microbenchmarks and image classification with ResNet-50. Our algorithms allow us to prototype a model for a mesh-tangling dataset, where sample sizes are very large. We show that our parallelization achieves excellent strong and weak scaling and enables training for previously unreachable datasets.
AB - Scaling CNN training is necessary to keep up with growing datasets and reduce training time. We also see an emerging need to handle datasets with very large samples, where memory requirements for training are large. Existing training frameworks use a data-parallel approach that partitions samples within a mini-batch, but limits to scaling the minibatch size and memory consumption makes this untenable for large samples. We describe and implement new approaches to convolution, which parallelize using spatial decomposition or a combination of sample and spatial decomposition. This introduces many performance knobs for a network, so we develop a performance model for CNNs and present a method for using it to automatically determine efficient parallelization strategies. We evaluate our algorithms with microbenchmarks and image classification with ResNet-50. Our algorithms allow us to prototype a model for a mesh-tangling dataset, where sample sizes are very large. We show that our parallelization achieves excellent strong and weak scaling and enables training for previously unreachable datasets.
KW - Algorithms
KW - Convolution
KW - Deep learning
KW - HPC
KW - Performance modeling
UR - http://www.scopus.com/inward/record.url?scp=85072053704&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85072053704&partnerID=8YFLogxK
U2 - 10.1109/IPDPS.2019.00031
DO - 10.1109/IPDPS.2019.00031
M3 - Conference contribution
AN - SCOPUS:85072053704
T3 - Proceedings - 2019 IEEE 33rd International Parallel and Distributed Processing Symposium, IPDPS 2019
SP - 210
EP - 220
BT - Proceedings - 2019 IEEE 33rd International Parallel and Distributed Processing Symposium, IPDPS 2019
PB - Institute of Electrical and Electronics Engineers Inc.
ER -