Infrastructure-level Support for GPU-Enabled Deep Learning in DATAVIEW

Scientific workflow has become a common practice for scientists to effectively formalize and structure complex scientific processes, which in turn has accelerated scientific discoveries in numerous research fields. With the recent thriving of deep learning in broad range of scientific projects, there is a rising need for deep learning support in scientific workflow infrastructures — SWFMSs. However, current GPU-enabled deep learning frameworks are developed separately, not suitable for direct exploitation in SWFMSs, which forces scientists to handle deep learning outside of SWFMSs and then integrate in workflows in an ad-hoc manner. What workflow users pressingly need today is a user-friendly and well-integrated SWFMS to facilitate GPU-enabled deep learning as native workflows so that they can conveniently design, train, reuse, and integrate deep learning models in comprehensive workflows. In this paper, We report our latest research progress in supporting GPU-enabled deep learning at infrastructure-level in a popular SWFMS — DATAVIEW, which facilitates: (1) fast design, train, reuse neural networks as native workflows per Deep-Learning-as-a- Workflow (DLaaW) or integrate pre-trained neural network models with ordinary Tasks in one comprehensive workflow via JAVA API or WebBench GUI; (2) flexibly leverage various types of GPU resources for executing deep learning workflows. Our approach and implementations are thoroughly evaluated through experiments that demonstrate the efficacy and efficiency as compared to conventional Pytorch-based implementations.