Kernel- vs. User-Level Networking: A Ballad of Interrupts and How to Mitigate Them

Abstract

Networking performance has become especially important in the current age with growing demands on services over the Internet. Recent advances in network controllers has exposed bottlenecks in various parts of network processing. User-level networking, which bypasses the operating system's network stack and replaces it with one re-implemented in the userspace, is often framed as a silver bullet to mitigate any performance issues arising in the kernel network stack. However, there is often no comprehensive study on where this performance increase ultimately comes from.

This work aims to explore potential areas from which improvements in overall performance can arise. Most importantly, it is identified that asynchronous interrupts and their handling is a major source of overhead associated with the kernel network stack. Several proposals are presented with the goal of reducing the need for interrupts in the kernel network stack, simulating the execution model of user-level networking. It is shown that a small kernel modification with around 30 lines of code change results in a substantial performance increase without the need to replace the kernel network stack in its entirety.