| Description |
Containers are an increasingly popular packaging framework for complex applications, offering benefits such as lightweight isolation, portability, and ease of deployment. These benefits have the potential to solve a myriad of issues that have long been present in the high performance computing world, presenting a compelling narrative for their use in these environments. Unfortunately, standard container runtimes rely on relatively slow, virtualized network stacks to achieve network isolation and portability, which are incompatible with the kernel bypass networking technology RDMA. This has limited the widespread adoption of containers in high performance computing environments where RDMA-reliant applications are quite common. Recent projects such as FreeFlow and Mellanox's RDMA hardware have created solutions for enabling RDMA inside of containers while maintaining varying degrees of the benefits of standard containers. However, despite the strong claims made by these solutions, many of their characteristics such as performance, isolation sacrifices, and scalability are either not well documented or simply not known; these characteristics being critical to assess for several high performance computing use cases. This thesis attempts to remedy this issue by identifying, explaining, measuring, and comparing these characteristics for the various solutions available today for enabling RDMA in containers; ultimately providing high performance computing end-users a more holistic perspective on which solutions may be most viable for their environment and use case. |
