Tools
Tools
Krasniqi, Astrit, Salekzamankhani, Shahram and Virdee, Bal (2026) Last-mile queueing and HTTP/3 performance: a large-scale empirical evaluation on real access networks. Computer Networks. p. 112722. ISSN 13891286
This paper presents a large-scale empirical evaluation of HTTP/3 (QUIC) and HTTP/2 performance in real last-mile access networks. Unlike prior work that relies heavily on cloud-hosted virtual machines or simulated testbeds, our measurements were collected using physical routers, Ethernet and Wi-Fi links, and consumer-grade access equipment representative of typical small office/home office (SOHO) deployments. More than half a million end-to-end RTT samples were gathered under baseline conditions, controlled packet loss, uplink saturation, and combined delay-plus-loss scenarios.
Across both wired and wireless paths, HTTP/3 and HTTP/2 exhibit comparable median latencies. However, their tail behaviour diverges significantly when the access link is congested. Under a competing TCP upload, HTTP/3 experiences elevated 95th–99th percentile delays and occasional timeouts, consistent with mixed TCP/UDP queueing unfairness at the consumer access link. Consumer NAT and state-handling behaviour may contribute further pressure, but were not directly instrumented in this study. When the competing flow also uses QUIC, HTTP/3 tail performance stabilises and can match or exceed HTTP/2.
These results show that HTTP/3 performance is shaped more by last-mile queueing behaviour than by the transport protocol itself. In typical SOHO environments, switching between HTTP/2 and HTTP/3 does not guarantee lower latency; instead, performance depends primarily on access-link buffering, uplink load, and queue-management discipline. This study highlights the need to evaluate transport innovations in realistic access-network conditions to understand their practical behaviour.
Available under License Creative Commons Attribution 4.0.
Download (3MB) | Preview
 |
View Item |