An empirical study on program failures of deep learning jobs

Deep learning has made significant achievements in many application areas. To train and test models more efficiently, enterprise developers submit and run their deep learning programs on a shared, multi-tenant platform. However, some of the programs fail after a long execution time due to code/script defects, which reduces the development productivity and wastes expensive resources such as GPU, storage, and network I/O. This paper presents the first comprehensive empirical study on program failures of deep learning jobs. 4960 real failures are collected from a deep learning platform in Microsoft. We manually examine their failure messages and classify them into 20 categories. In addition, we identify the common root causes and bug-fix solutions on a sample of 400 failures. To better understand the current testing and debugging practices for deep learning, we also conduct developer interviews. Our major findings include: (1) 48.0% of the failures occur in the interaction with the platform rather than in the execution of code logic, mostly due to the discrepancies between local and platform execution environments; (2) Deep learning specific failures (13.5%) are mainly caused by inappropriate model parameters/structures and framework API misunderstanding; (3) Current debugging practices are not efficient for fault localization in many cases, and developers need more deep learning specific tools. Based on our findings, we further suggest possible research topics and tooling support that could facilitate future deep learning development. CCS CONCEPTS • Software and its engineering → Software defect analysis.