Doug Jacobsen

Container-based computed is rapidly changing the way software is developed, tested, and deployed. This paper builds on previously presented work on a prototype framework for running containers on HPC platforms. We will present a detailed overview of the design and implementation of Shifter, which in partnership with Cray has extended on the early prototype concepts and is now in production at NERSC. Shifter enables end users to execute containers using images constructed from various methods including the popular Docker-based ecosystem. We will discuss some of the improvements over the initial prototype including an improved image manager, integration with SLURM, integration with the burst buffer, and user controllable volume mounts. In addition, we will discuss lessons learned, performance results, and real-world use cases of Shifter in action. We will also discuss the potential role of containers in scientific and technical computing including how they complement the scientific process. We will conclude with a discussion about the future directions of Shifter.