Making QUIC Quicker With NIC Offload

This paper aims at defining the right set of primitives a NIC shall expose to efficiently offload the QUIC protocol. Although previous work already partially tackled this problem, it has only considered one specific aspect: the crypto module. We instead dissect different QUIC implementations, and perform an in-depth analysis of the cost associated to many of its components. We find that the kernel to userspace communication, the crypto module and the packet reordering algorithm are CPU hungry and often the cause of application performance degradation. We use those findings to define an architecture for offloading QUIC and discuss the associated challenges.