A Comprehensive Cloud Architecture for Machine Learning-enabled Research

The success of machine learning (ML) algorithms, and deep learning in particular, is having a transformative impact on a wide range of research disciplines, from astronomy, materials science, and climate change to bioinformatics, computational health, and animal ecology. At the same time, these new techniques introduce computational modalities that create challenges for academic computing centers and resource providers that have historically focused on asynchronous, batch-computing paradigms. In particular, there is an emergent need for computing models that enable efficient use of specialized hardware such as graphical processing units (GPUs) in the presence of interactive workloads. In this paper, we present a comprehensive, cloud-based architecture comprised of open-source software layers to better meet the needs of modern ML processes and workloads. This framework, deployed at the Texas Advanced Computing Center and in use by various research teams, provides different interfaces at varying levels of abstraction to support and simplify the tasks of users with different backgrounds and expertise, and to efficiently leverage limited GPU resources for these tasks. We present techniques and implementation details for overcoming challenges related to developing and maintaining such an infrastructure which will be of interest to service providers and infrastructure developers alike.