Kernel Bypass Surgery: A Viable Procedure for Maximizing QUIC Bandwidth?

Authors: Johannes Späth, Stefan Lachnit, Marcel Kempf, Kilian Holzinger, Georg Carle, Johannes Zirngibl

Published in IEEE Network Operations and Management Symposium (NOMS), 2026

Abstract:
QUIC is a protocol that aims to solve many problems of the TCP/TLS stack. Previous work has shown that the performance of those implementations differs greatly in high-bandwidth scenarios. One major reason for this is that QUIC is usually implemented in user space, building on the existing UDP functionality offered by the kernel network stack. While this design decision enables fast development cycles, it comes at the cost of a high number of context switches between kernel and user space. In this paper, we investigate the impact of kernel bypass on performance. We integrated the Data Plane Development Kit (DPDK) into three QUIC implementations, bypassing their use of the kernel networking stack, and updated a fourth stack that previous work combined with DPDK. We compare two deployment scenarios for our DPDK QUIC implementations: (1) running exclusively on the NIC port, and (2) sharing a port with the kernel using Single Root I/O Virtualization (SR-IOV). Our analysis shows that kernel bypass can greatly increase the goodput achieved with QUIC, with a speedup reaching up to a factor of 3× and goodput over 10 GBit/s. However, the reachable performance increase highly depends on the implementation, and in one case performance did also not further increase when using DPDK. Furthermore, we show that offloading techniques can reach similar performance if offered by the system and used correctly by the implementations.

Recommended citation: Johannes Späth, Stefan Lachnit, Marcel Kempf, Kilian Holzinger, Georg Carle, Johannes Zirngibl, "Kernel Bypass Surgery: A Viable Procedure for Maximizing QUIC Bandwidth?." IEEE Network Operations and Management Symposium (NOMS), 2026.