Synergistic Server-Based Network Processing Stack
Published in Ph.D. defense - University of Colorado Boulder, 2022
Network functions provide the required functionality to interconnect systems while ensuring security, availability, efficiency, and performance. With the recent trend to run network functions in software over commodity servers (instead of using specialized appliances), there is the need to introduce new systems that can provide the required network features while running on top of optimal processing environments. Several software packet processing technologies currently exist (e.g., in-kernel/kernel-bypass, XDP, and SmartNICs) and each of them provide different features in terms of available functionality, processing capabilities, performance, and efficiency. In this thesis, we break down network application processing needs and by characterizing the processing features provided by each technology, we verify that no single technology can cover all network application requirements. With this observation, we provide new systems that often break the boundaries between different technologies, allowing the building of optimal packet processing environments that can meet countless requirements of modern networks. Using this as the foundation of our work, we build systems to address many network function needs – layer 2 to layer 7 processing and monitoring. In our first work, we introduce a new packet I/O subsystem to a high-performance userspace TCP stack. This subsystem is provided by new programmable in-kernel features allowing the TCP stack to have a better resource consumption profile and to build cooperation mechanisms between the kernel and userspace. In our second work, we address the needs of monitoring systems by introducing new primitives that allow for building high-coverage monitoring systems with high performance and efficiency. We optimize those primitives by building an efficient division of work between SmartNIC offloads, XDP on the host, and userspace processing. Finally, in our third work, we rethink the Linux networking stack to address the inneficiencies that prevent it to support the performance requirements of modern applications. We propose to break down its processing in a minimal and efficient fast path and a in a robust and feature-rich slow path provided by the Linux kernel. The fast path is built on demand, based on current processing needs for a given set of services configuration and gets assistance from the slow path for processing completeness. This allows avoiding unnecessary processing inside the kernel, minimizing overheads and increasing performance, while still maintaining Linux’s rich set of features. With these contributions, we believe that we can provide new foundations to help both academia and industry to build optimized systems that can address many modern network needs.