Integrated Model, Batch and Domain Parallelism in Training Neural Networks
- Lawrence Berkeley National Lab. (LBNL), Berkeley, CA (United States)
We propose a new integrated method of exploiting model, batch and domain parallelism for the training of deep neural networks (DNNs) on large distributed-memory computers using minibatch stochastic gradient descent (SGD). Our goal is to find an efficient parallelization strategy for a fixed batch size using $$P$$ processes. Our method is inspired by the communication-avoiding algorithms in numerical linear algebra. We see $$P$$ processes as logically divided into a $$P_r \times P_c$$ grid where the $$P_r$$ dimension is implicitly responsible for model/domain parallelism and the $$P_c$$ dimension is implicitly responsible for batch parallelism. In practice, the integrated matrix-based parallel algorithm encapsulates these types of parallelism automatically. We analyze the communication complexity and analytically demonstrate that the lowest communication costs are often achieved neither with pure model nor with pure data parallelism. We also show how the domain parallel approach can help in extending the theoretical scaling limit of the typical batch parallel method.
- Research Organization:
- Lawrence Berkeley National Laboratory (LBNL), Berkeley, CA (United States)
- Sponsoring Organization:
- USDOE Office of Science (SC), Advanced Scientific Computing Research (ASCR) (SC-21)
- DOE Contract Number:
- AC02-05CH11231
- OSTI ID:
- 1454501
- Country of Publication:
- United States
- Language:
- English
Similar Records
Train Like a (Var)Pro: Efficient Training of Neural Networks with Variable Projection
Improving scalability of parallel CNN training by adaptively adjusting parameter update frequency