Near-Optimal Sparse Allreduce for Distributed Deep Learning

TL;DR

This paper introduces O$k$-Top$k$, a scalable sparse allreduce algorithm for distributed deep learning that reduces communication overhead and maintains model accuracy, significantly improving training throughput on large GPU clusters.

Contribution

It presents a novel sparse allreduce algorithm with asymptotically optimal communication volume integrated with decentralized SGD, and demonstrates its effectiveness in large-scale training.

Findings

Achieves similar accuracy to dense allreduce.

Significantly improves training throughput (up to 12.95x on BERT).

More scalable than existing methods.

Abstract

Communication overhead is one of the major obstacles to train large deep learning models at scale. Gradient sparsification is a promising technique to reduce the communication volume. However, it is very challenging to obtain real performance improvement because of (1) the difficulty of achieving an scalable and efficient sparse allreduce algorithm and (2) the sparsification overhead. This paper proposes O$k$-Top$k$, a scheme for distributed training with sparse gradients. O$k$-Top$k$ integrates a novel sparse allreduce algorithm (less than 6$k$ communication volume which is asymptotically optimal) with the decentralized parallel Stochastic Gradient Descent (SGD) optimizer, and its convergence is proved. To reduce the sparsification overhead, O$k$-Top$k$ efficiently selects the top-$k$ gradient values according to an estimated threshold. Evaluations are conducted on the Piz Daint supercomputer with neural network models from different deep learning domains. Empirical results show that O$k$-Top$k$ achieves similar model accuracy to dense allreduce. Compared with the optimized dense and the state-of-the-art sparse allreduces, O$k$-Top$k$ is more scalable and significantly improves training throughput (e.g., 3.29x-12.95x improvement for BERT on 256 GPUs).