Trade-offs of Local SGD at Scale: An Empirical Study
This work addresses the scalability challenges in distributed deep learning for large datasets and models, though it is incremental in highlighting limitations at scale.
The study tackled the trade-offs of using local SGD for distributed training at scale, finding that while it reduces communication costs and speeds up training, it leads to lower accuracy compared to smaller-scale experiments, with slow momentum improving accuracy without extra communication.
As datasets and models become increasingly large, distributed training has become a necessary component to allow deep neural networks to train in reasonable amounts of time. However, distributed training can have substantial communication overhead that hinders its scalability. One strategy for reducing this overhead is to perform multiple unsynchronized SGD steps independently on each worker between synchronization steps, a technique known as local SGD. We conduct a comprehensive empirical study of local SGD and related methods on a large-scale image classification task. We find that performing local SGD comes at a price: lower communication costs (and thereby faster training) are accompanied by lower accuracy. This finding is in contrast from the smaller-scale experiments in prior work, suggesting that local SGD encounters challenges at scale. We further show that incorporating the slow momentum framework of Wang et al. (2020) consistently improves accuracy without requiring additional communication, hinting at future directions for potentially escaping this trade-off.