Joint Dynamic Data and Model Parallelism for Distributed Training of DNNs Over Heterogeneous Infrastructure

Abstract

Distributed training of deep neural networks (DNNs) suffers from efficiency declines in dynamic heterogeneous environments, due to the resource wastage brought by the straggler problem in data parallelism (DP) and pipeline bubbles in model parallelism (MP). Additionally, the limited resource availability requires a trade-off between training performance and long-term costs, particularly in online settings. To address these challenges, this article presents a novel online approach to maximize long-term training efficiency in heterogeneous environments through uneven data assignment and communication-aware model partitioning. A group-based hierarchical architecture combining DP and MP is developed to balance discrepant computation and communication capabilities, and offer a flexible parallel mechanism. In order to jointly optimize the performance and long-term cost of the online DL training process, we formulate this problem as a stochastic optimization with time-averaged constraints. By utilizing Lyapunov’s stochastic network optimization theory, we decompose it into several instantaneous sub-optimizations, and devise an effective online solution to address them based on tentative searching and linear solving. We have implemented a prototype system and evaluated the effectiveness of our solution based on realistic experiments, reducing batch training time by up to 68.59% over state-of-the-art methods.