Containerized AI Framework on Secure Shared Multi-GPU Systems

Nowadays, technology has reached the point where Graphics Processing Unit (GPU) drive Artificial Intelligence (AI) workloads with large datasets to speed up the process of solving more complex real-world problems. The DevOps movement has brought container-based deployment models into the world of software development and application deployment. Data scientists can quickly prototype, build, and deploy a containerized application or service, allowing them to manage AI workloads more effectively with portability. Docker is a widely used containerization platform, requiring root privileges for execution, which has associated security concerns and may not be readily adopted in secure shared multi-GPU systems. Singularity addresses the needs of container-based virtualization in a secure environment, which allows the same Docker containers to be used without any changes or security concerns. This work presents the challenges and approaches for quickly deploying complex AI workloads on multi-GPU systems without compromising security by ensuring that AI frameworks run in an isolated container environment with user credentials. In addition, a comprehensive performance evaluation of containers converted from Docker to Singularity in the context of native is performed on a multi-GPU system to solve more complex real-world problems with AI workflow containerization.