PipeMare: Asynchronous Pipeline Parallel DNN Training
This addresses the problem of inefficient hardware utilization in pipeline parallelism for deep learning practitioners, offering a novel solution that is not incremental but provides specific gains.
The paper tackled the trade-off between hardware efficiency and statistical efficiency in pipeline parallelism for DNN training by introducing PipeMare, a method that tolerates asynchronous updates, resulting in up to 2.7x less memory usage or 4.3x higher pipeline utilization with similar model quality compared to state-of-the-art synchronous techniques.
Pipeline parallelism (PP) when training neural networks enables larger models to be partitioned spatially, leading to both lower network communication and overall higher hardware utilization. Unfortunately, to preserve the statistical efficiency of sequential training, existing PP techniques sacrifice hardware efficiency by decreasing pipeline utilization or incurring extra memory costs. In this paper, we investigate to what extent these sacrifices are necessary. We devise PipeMare, a simple yet robust training method that tolerates asynchronous updates during PP execution without sacrificing utilization or memory, which allows efficient use of fine-grained pipeline parallelism. Concretely, when tested on ResNet and Transformer networks, asynchrony enables PipeMare to use up to $2.7\times$ less memory or get $4.3\times$ higher pipeline utilization, with similar model quality, when compared to state-of-the-art synchronous PP training techniques.