Accelerating Collective Communication in Data Parallel Training across Deep Learning Frameworks

Explore a 17-minute conference talk from USENIX NSDI '22 that delves into accelerating collective communication in data parallel training across deep learning frameworks. Learn about new techniques developed within Horovod, a generic communication library, to improve the control plane and enhance performance in large-scale distributed training. Discover how the researchers implemented a caching strategy and decentralized orchestration to optimize the coordinator-worker logic, and introduced a feature for users to group collective operations for finer control over communication buffer sizes. Examine the experimental results conducted on the Summit supercomputer, comparing the proposed strategies against Horovod's original design, tf.distribute, torch.DDP, and BytePS. Gain insights into the impressive performance improvements achieved, including a 2x speedup at 6000 GPUs scale and near-linear scaling of 0.93 with 1.54 exaflops sustained performance using 27,600 GPUs on a scientific application (STEMDL).