Fine-grained OS Control over High-performance Networking

Research output: Types of thesisDoctoral thesis

Contributors

Abstract

In the evolving public cloud environment, RDMA (Remote Direct Memory Access) networks surpass traditional TCP/IP in performance but face integration hurdles across applications, hardware, and network stacks. This complexity has even prompted cloud providers to develop custom NICs and protocols for a better performance-usability balance. This thesis focuses on how operating system (OS) techniques can enhance RDMA network usability with minimal performance trade-offs.

A significant hurdle in the usability of RDMA networks is kernel bypass, which removes the OS from the communication dataplane. While effective in High-Performance Computing, kernel bypass is problematic in cloud environments, where diverse applications rely on OS support for resource sharing. Removing kernel bypass, as is done in TCP/IP, is not an option, as it would nullify the performance benefits of RDMA. Therefore, the challenge is to make the RDMA network dataplane more controllable with minimal changes to the existing RDMA network stack.

To address this, the OS can interpose on the RDMA dataplane continuously or intermittently to enhance overall control over RDMA communication. Continuous interposition redirects the dataplane through the OS kernel, offering control with minimal impact, enabling tasks like performance monitoring and rate limiting. Intermittent interposition, ideal for scenarios demanding low overhead, alternates communication between a high-performance bypass and an OS-managed manipulation phase, allowing for the implementation of, for example, transparent live migration for RDMA applications. These methods, continuous and intermittent interposition, significantly improve OS control over the RDMA dataplane, facilitating wider adoption and commoditization of RDMA networks in cloud infrastructures.

Details

Original languageEnglish
Qualification levelDr.-Ing.
Awarding Institution
Supervisors/Advisors
  • Härtig, Hermann, Main supervisor
  • Fetzer, Christof, Supervisor
  • Brinkmann, André, Reviewer, External person
Thesis sponsors
  • Barkhausen Institut
Defense Date (Date of certificate)4 Sept 2024
Publication statusPublished - 2024
No renderer: customAssociatesEventsRenderPortal,dk.atira.pure.api.shared.model.researchoutput.Thesis