Taming unbalanced training workloads in deep learning with partial collective operations
Open accessThe DOI matches its Crossref record; 67 of the 67 references it cites were found in a registry or an index.
Abstract
Load imbalance pervasively exists in distributed deep learning training systems, either caused by the inherent imbalance in learned tasks or by the system itself. Traditional synchronous Stochastic Gradient Descent (SGD) achieves good accuracy for a wide variety of tasks, but relies on global synchronization to accumulate the gradients at every training step. In this paper, we propose eager-SGD, which relaxes the global synchronization for decentralized accumulation. To implement eager-SGD, we propose to use two partial collectives: solo and majority. With solo allreduce, the faster processes contribute their gradients eagerly without waiting for the slower processes, whereas with majority allreduce, at least half of the participants must contribute gradients before continuing, all without using a central parameter server. We theoretically prove the convergence of the algorithms and describe the partial collectives in detail. Experiments are conducted on a variety of neural networks and datasets. The results on load-imbalanced environments show that eager-SGD achieves 2.64 X speedup (ResNet-50 on ImageNet) over the asynchronous centralized SGD, and achieves 1.29 X speedup (ResNet-50 on ImageNet) and 1.27X speedup (LSTM on UCF101) over the state-of-the-art synchronous decentralized SGDs, without losing accuracy.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 3×
“Residual Network (ResNet) [21] is widely used in computer vision tasks.”
MethodVaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1×
“Transformers [57] are sequence-to-sequence models that translate a sequence of words from one language to another.”
BackgroundVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×
“Early convolutional networks demonstrated groundbreaking successes in computer vision, ranging from image classification to object detection [29, 49].”
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 1×
“Long short-term memory (LSTM) [23] is a type of unit cell in Recurrent Neural Networks (RNN).”
BackgroundLeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
Exploiting Generative AI to Scale up Intelligent Tutoring Systems
Jakubův, J., Chvalovský, K., Goertzel, Z., Kaliszyk, C., Olšák, M., Piotrowski, B., Schulz, S., Suda, M., & Urban, J. (2023) · DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) · doi:10.4230/lipics.itp.2023.19 · Cited by 76,658
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173 · cited 1×
“deoclassification,naturallanguageprocessingfor machinetranslation,questionanswering,textcomprehension,andsynthetictextgeneration.Thelattermodels containmorethan1.5billionparametersandtakeweeks totrain[14,42].Otherdemandingneuralnetworksare trainedonthelargestsupercomputerstoachievescientificbreakthroughs[33,38].Furthe…”
ImageNet: A large-scale hierarchical image database
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333 · cited 1×
“4 shows the runtime distribution over the sampled batches for 5 epochs of training for the classic ResNet-50 model [21] on ImageNet [14], on a standard Google Cloud instance (n1-standard-16 with 2x Nvidia V100 GPUs).”
MethodGradient-based learning applied to document recognition
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998) · Proceedings of the IEEE · doi:10.1109/5.726791 · Cited by 58,805 · cited 1×
“The gradient of those functions can be computed using the backpropagation algorithm [35], processing the nodes in the DAG in a reverse topological order.”
Densely Connected Convolutional Networks
Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.243 · Cited by 48,566
Research works printed but not linked: 15
- Alistarh Dan, Advances in Neural Information Processing Systems 31. Curran Associates
- Chilimbi Trishul, 2014, 11th USENIX Symposium on Operating Systems Design and Implementation (OSDI 14)
- Dean Jeffrey, 2012, Ng
- Devries Terrance, 2017, Taylor
- Donahue Jeff, 2014, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell.
- Gibiansky Andrew, 2017, Bringing HPC techniques to deep learning. (2017). URL http://research.baidu.com/bringing-hpc-techniques-deep-learning
- Hoefler T., International Conference for High Performance Computing, Networking, Storage and Analysis (SC'10)
- Hsieh Kevin, 2017, Proceedings of the 14th USENIX Conference on Networked Systems Design and Implementation (NSDI'17)
- Jayarajan Anand, 2019, Proceedings of the 2nd SysML Conference.
- 10.5555/3295222.3295285 · doi:10.5555/3295222.3295285
- Lian Xiangru, 2018, Proceedings of the 35th International Conference on Machine Learning (Proceedings of Machine Learning Research), Jennifer Dy and Andreas Krause (Eds.), 80
- Interface Forum Message Passing, 2015, MPI: A Message-Passing Interface Standard Version 3.1.
- Soomro Khurram, 2012, Amir Roshan Zamir, and Mubarak Shah
- 10.5555/3020948.3021030 · doi:10.5555/3020948.3021030
- 10.1109/cvpr.2015.7299101 · DOI · Paper page
Cited by
Energy efficiency in cloud computing data centers: a survey on software technologies
Katal, A., Dahiya, S., & Choudhury, T. (2023) · Cluster Computing · doi:10.1007/s10586-022-03713-0 · Cited by 516
“…techniques there are various other supervised learning techniques like Support Vector Machine (SVM) [103], Stochastic Gradient Descent (SGD) [104], Logistic Regression (LR) [105], Random Forest (RF) [106], Multi-Layer Perceptron (MLP) [107], Backpropagation neural network [108] were used.”
MethodAn In-Depth Analysis of the Slingshot Interconnect
Sensi, D. D., Girolamo, S. D., McMahon, K. H., Roweth, D., & Hoefler, T. (2020) · International Conference for High Performance Computing, Networking, Storage and Analysis (SC) · doi:10.1109/sc41405.2020.00039 · Cited by 154
“Resnet-proxy This is a ML/AI proxy application [41], reproducing the communication phases of a Deep500 benchmark [42] Residual Neural Network (resnet).”
MethodSelf-citationChimera: efficiently training large-scale neural networks with bidirectional pipelines
Li, S., & Hoefler, T. (2021) · International Conference for High Performance Computing, Networking, Storage and Analysis (SC) · doi:10.1145/3458817.3476145 · Cited by 223
“To scale up the training process to parallel machines, data parallelism [20, 33, 49, 60] is the common method, in which the minibatch is partitioned among P workers and each worker maintains a copy of the entire model.”
MethodSelf-citationFasterMoE: modeling and optimizing training of large-scale dynamic pre-trained models
He, J., Zhai, J., Antunes, T., Wang, H., Luo, F., Shi, S., & Li, Q. (2022) · ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming · doi:10.1145/3503221.3508418 · Cited by 254
“Asynchronous partial model updating methods [15, 16, 18, 32] are introduced to speed up convergence in heterogeneous environment for data parallelism.”
MethodParallel and Distributed Graph Neural Networks: An In-Depth Concurrency Analysis
Besta, M., & Hoefler, T. (2024) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2023.3303431 · Cited by 106
“Thismaycause load imbalance ), similarly to, e.g., videos [16].”
Self-citation
Metadata
- Authors
- Li, S., Ben-Nun, T., Girolamo, S. D., Alistarh, D., & Hoefler, T.
- Affiliations
- ETH Zurich (CH)
- ETH Zurich
- Institute of Science and Technology Austria (AT)
- Venue
- Proceedings of the 25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming
- Publisher
- ACM
- Published
- 2020-02-19
- Pages
- 45-61
- OpenAlex
- W2967558351
- Cited by
- 57
- References
- 67
- Citations by year
- 2026: 2 · 2025: 10 · 2024: 7 · 2023: 8 · 2022: 11 · 2021: 11 · 2020: 8
- FWCI
- 3.63
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- European Research Council (ERC) under the European Union?s Horizon 2020 programme, grant agreement DAPP, (No. 678880)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Stochastic Gradient Optimization Techniques
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
- Keywords
- load imbalance, distributed deep learning training, AllReduce, asynchronous SGD, convergence analysis, speedup
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Taming unbalanced training workloads in deep learning with partial collective operations. https://app.cochat.ai/papers/10.1145/3332466.3374528
- The paper
- Li, S., Ben-Nun, T., Girolamo, S. D., Alistarh, D., & Hoefler, T. (2020). Taming unbalanced training workloads in deep learning with partial collective operations. Proceedings of the 25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming. https://doi.org/10.1145/3332466.3374528
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.