Understanding QUIC's Throughput Speedbumps

QUIC is experiencing phenomenal success as a low-latency, secure, and flexible protocol, adopted and deployed by multiple hyperscale content providers. However, the throughput of all its implementations remains remarkably low compared to that of TCP + TLS. Approaches to accelerating QUIC's throughput remain scattered, applied to a single implementation at a time, and one approach at a time. In this paper, we systematically identify the different bottlenecks that arise in different implementations. We use our findings to propose an application-agnostic architecture to accelerate single-connection throughput in QUIC. We demonstrate the value of our proposal by accelerating two implementations: quicly and mvfst, improving their throughput by 1.36× and 1.88× respectively, compared to their baseline implementations.