DLHub: Model and Data Serving for Science

Abstract--While the Machine Learning (ML) landscape is evolving rapidly, there has been a relative lag in the development of the "learning systems" needed to enable broad adoption. Furthermore, few such systems are designed to support the specialized requirements of scientific ML. Here we present the Data and Learning Hub for science (DLHub), a multi-tenant system that provides both model repository and serving capabilities witha focus on science applications. First, its selfservice modelrepository allows users to share, publish, verify, reproduce, and reuse models, and addresses concerns related to model reproducibility by packaging and distributing models and all constituent components. Second, it implements scalable and low-latency serving capabilities that can leverage parallel and distributed computing resources to democratize access to published modelsthrough a simple web interface. Unlike other model serving frameworks, DLHub can store and serve any Python 3-compatible model or processing function, plus multiple-function pipelines. We show that relative to other model serving systems including TensorFlow Serving, SageMaker, and Clipper, DLHub provides greater capabilities, comparable performance without memoization and batching, and significantly better performance when the latter two techniques can be employed. We also describe early uses of DLHub for scientific applications. I. INTRODUCTION Machine Learning (ML) is disrupting nearly every aspect of computing. Researchers now turn to ML methods to uncover patterns in vast data collections and to make decisions with little or no human input. As ML becomes increasingly pervasive, newsystems are required to support the development, adoption, and application of ML. We refer to the broad class of systems designed to support ML as "learning systems." Learning systems need to support the entire ML lifecycle (see Figure 1), including model development [1, 2]; scalable training across potentially tens of thousands of cores and GPUs [3]; model publication and sharing [4]; and low latency and highthroughput inference[5]; all while encouraging best-practice software engineering when developing models [6].