arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2201.07598cs.DCcs.LG
收藏

面向分布式深度学习的近最优稀疏 Allreduce

Near-Optimal Sparse Allreduce for Distributed Deep Learning

Department of Computer Science, ETH Zurich(计算机科学系,苏黎世联邦理工学院)

Shigang Li, Torsten Hoefler

AI总结:

本文提出 O$k$-Top$k$ 方案,通过渐近最优的稀疏 allreduce 算法与去中心化 SGD 结合,并利用阈值估计高效选择 top-$k$ 梯度,在保持模型精度的同时显著提升分布式深度学习训练吞吐量和可扩展性。

摘要:

通信开销是在大规模环境下训练大型深度学习模型的主要障碍之一。梯度稀疏化是一种有前景的降低通信量的技术。然而,由于(1)难以实现可扩展且高效的稀疏 allreduce 算法,以及(2)稀疏化本身带来的开销,获得真实性能提升非常具有挑战性。本文提出 O$k$-Top$k$,一种用于稀疏梯度分布式训练的方案。O$k$-Top$k$ 将一种新颖的稀疏 allreduce 算法(通信量小于 6$k$,渐近最优)与去中心化并行随机梯度下降(SGD)优化器相结合,并证明了其收敛性。为降低稀疏化开销,O$k$-Top$k$ 根据估计的阈值高效地选择 top-$k$ 梯度值。在 Piz Daint 超级计算机上使用来自不同深度学习领域的神经网络模型进行了评估。实验结果表明,O$k$-Top$k$ 达到了与稠密 allreduce 相似的模型精度。与优化的稠密 allreduce 和最先进的稀疏 allreduce 相比,O$k$-Top$k$ 具有更好的可扩展性,并显著提高了训练吞吐量(例如,在 256 个 GPU 上训练 BERT 时提升 3.29 倍至 12.95 倍)。

英文摘要:

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).

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑