Demystifying Developers' Issues in Distributed Training of Deep Learning Software

Deep learning (DL) has been pervasive in a wide spectrum of nowadays software systems and applications. The rich features of these DL based software applications (i.e., DL software) usually rely on powerful DL models. To train powerful DL models with large datasets efficiently, it has been a common practice for developers to parallelize and distribute the computation and memory over multiple devices in the training process, which is known as distributed training. However, existing efforts in the software engineering (SE) research community mainly focus on issues in the general process of training DL models. In contrast, to the best of our knowledge, issues that developers encounter in distributed training have never been well studied. Given the surging importance of distributed training in the current practice of developing DL software, this paper fills in the knowledge gap and presents the first comprehensive study on developers' issues in distributed training. To this end, we extract and analyze 1,054 real-world developers' issues in distributed training from Stack Overflow and GitHub, two commonly used data sources for studying software issues. We construct a fine-grained taxonomy consisting of 30 categories regarding the fault symptoms and summarize common fix patterns for different symptoms. Based on the results, we suggest actionable implications and research avenues that can potentially facilitate the future development of distributed training.