PowerSGD: Practical Low-Rank Gradient Compression for Distributed Optimization
This addresses the communication overhead problem for distributed machine learning training, offering a practical solution with incremental improvements over existing compression methods.
The paper tackles the communication bottleneck in distributed optimization by proposing PowerSGD, a low-rank gradient compression method that compresses gradients rapidly, aggregates them efficiently, and achieves test performance comparable to SGD, with demonstrated wall-clock speedups for convolutional networks and LSTMs on common datasets.
We study gradient compression methods to alleviate the communication bottleneck in data-parallel distributed optimization. Despite the significant attention received, current compression schemes either do not scale well or fail to achieve the target test accuracy. We propose a new low-rank gradient compressor based on power iteration that can i) compress gradients rapidly, ii) efficiently aggregate the compressed gradients using all-reduce, and iii) achieve test performance on par with SGD. The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD with an optimized communication backend. We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets. Our code is available at https://github.com/epfml/powersgd.