Research map / Optimization / Optimizers
On the Acceleration of Deep Learning Model Parallelism With Staleness
An Xu, Zhouyuan Huo, Heng Huang
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DSP instead of FR | Model-parallel CNN training · CIFAR-10, ResNet164, K=4 | speedup: x1.7 -> x2.7 | Improved |
| Use DSP instead of FR | Model-parallel CNN training · CIFAR-100, ResNet1001, K=4 | speedup: x1.9 -> x4.8 | Improved |
| Use DSP instead of FR | Model-parallel CNN training · ImageNet, ResNet50, K=3 | speedup: x1.6 -> x3.0 | Improved |
| Use DSP instead of FR | ResNet164 · CIFAR-10, K=3 | Top-1 test accuracy: 94.55% -> 94.68% | Improved |
| Use DSP instead of FR | ResNet164 · CIFAR-100, K=3 | Top-1 test accuracy: 76.25% -> 76.29% | Improved |
| Use DSP instead of BP | ResNet50 · ImageNet, K=3 | Top-1 test accuracy: 75.35% -> 74.91% | Hurt |
| Increase Layer-wise Staleness from DSP(1,1,0;4,2,0) to DSP(3,3,0;10,5,0) (ablation) | DSP on ResNet164 · CIFAR-10, K=3 | Top-1 test accuracy: 94.68% -> 93.37% | Hurt |
| Increase Layer-wise Staleness from DSP(1,1,0;4,2,0) to DSP(3,3,0;10,5,0) (ablation) | DSP on ResNet164 · CIFAR-100, K=3 | Top-1 test accuracy: 76.05% -> 76.29% | Mixed |
| Increase Layer-wise Staleness from DSP(1,1,0;4,2,0) to DSP(2,2,0;6,3,0) (ablation) | DSP on ResNet164 · CIFAR-10, K=3 | Top-1 test accuracy: 94.68% -> 93.98% | Hurt |
| Increase Layer-wise Staleness from DSP(1,1,1,0;6,4,2,0) to DSP(3,3,3,0;15,10,5,0) (ablation) | DSP on ResNet164 · CIFAR-10, K=4 | Top-1 test accuracy: 94.32% -> 93.34% | Hurt |
| Increase queue lengths / Layer-wise Staleness (ablation) | DSP with randomly slowed GPUs · ResNet164, CIFAR-10, K=3, GPU slowdown 150% | slow down percentage: 70.99% -> 53.42% | Improved |
| Use DSP instead of FR under random GPU slowdown | ResNet164 · CIFAR-10, K=3, GPU slowdown 150% | slow down percentage: 359.2% -> 70.99% | Improved |
Closest papers by meaning
- SAPipe: Staleness-Aware Pipeline for Data Parallel DNN Training (NeurIPS 2022 · similarity 0.90)
- Accumulated Decoupled Learning with Gradient Staleness Mitigation for Convolutional Neural Networks (ICML 2021 · similarity 0.87)
- Improving the Robustness of Deep Neural Networks via Stability Training (CVPR 2016 · similarity 0.87)
- Exploring Hidden Dimensions in Accelerating Convolutional Neural Networks (ICML 2018 · similarity 0.86)
- PETRA: Parallel End-to-end Training with Reversible Architectures (ICLR 2025 · similarity 0.85)
- Decoupled Parallel Backpropagation with Convergence Guarantee (ICML 2018 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
