Waiting for QUIC: Passive Measurements to Understand QUIC Deployments

TL;DR

The paper tackles the challenge of understanding real-world QUIC deployments by major hypergiants, where direct measurement is difficult due to business sensitivity. It introduces a non-intrusive passive measurement approach based on backscatter traffic from a network telescope, complemented by flow records and limited active probing to validate findings. Key contributions include mapping QUIC stack configurations, uncovering structured QUIC connection ID encodings used for load-balancer routing, detecting off-net deployments, and quantifying layer-7 load balancers across multiple providers over 2021–2025. The results demonstrate that passive observations can accurately reveal deployment characteristics, inform optimization and security considerations, and guide future measurement strategies in fast-evolving QUIC ecosystems.

Abstract

QUIC experiences a rapid adoption since its standardization in 2021, and hypergiants configure their infrastructure to optimize for QUIC performance. In this paper, we introduce a passive measurement method to study both the progressive rollout and individual hypergiant configurations during the last five years. By analyzing backscatter traffic of the UCSD network telescope, we are able to make the following observations. First, Meta, Google, and Cloudflare configure significantly different maximal retransmission numbers and timeouts. Second, we can identify different off-net deployments of hypergiants, using packet features, such as QUIC connection IDs, packet coalescence, and packet lengths. Third, we observe changing hypergiant deployment configurations during our different measurement periods. Fourth, connection IDs can allow further insights into load balancer deployments, such as the number of servers. We bolster our results using two orthogonal measurements: passive recording of QUIC flows and active probing.

Figures (16)

• Figure 1: Connection establishment using QUIC. Client connection ID (C1) is consistently used during the connection establishment but the initial server ID (S1) can be replaced by the server with ID (S2).
• Figure 2: Illustration of a Meta /24 load balancer frontend cluster. Our method enables L7LB quantification.
• Figure 3: Overview of our measurement setup and data corpus.
• Figure 4: QUIC versions of clients and servers in one month of telescope traffic in 2021 to 2025, and flow records off the same months in 2024 and 2025. Colors included in the legend but invisible in the bars contribute $<$0.1 % of the traffic. After standardization of QUIC in May 2021, QUICv1 is rapidly adopted in 2022.
• Figure 4: $F_1$-score of SCID classifiers. Low $F_1$-scores for Google in 2022 originate from not using information encoding. Meta off-net classifiers consider more bits than Google classifiers and achieve better scores.