Cloudflare Tunes Congestion Control with Global Performance Data, ~10% Faster

Cloudflare has detailed a new performance initiative: using its daily traffic data, which spans essentially every major network on the planet, to tune congestion control algorithms for the specific characteristics of each network path and deliver responses faster. Early results show an average improvement of about 10%, currently running on free-tier QUIC traffic, with a plan to extend to TCP and all customers through 2026.

Why Congestion Control Matters

When packets travel from a client to a Cloudflare edge node, they cross many hops, and one of them is always the bottleneck of the path. If the sending rate exceeds the capacity of that bottleneck link, data is either queued (causing bufferbloat) or dropped, triggering TCP/QUIC retransmissions that can make congestion worse.

Traditional congestion control algorithms react to packet loss after the fact. Newer model-based algorithms such as BBR estimate bandwidth and round-trip time to send proactively. Cloudflare believes there is room to go further: existing algorithms only use information from the current connection, while Cloudflare can see the real-time state of the wider Internet.

Congestion control determines how much of the available bandwidth a user actually gets. On a mobile network with 200 ms RTT and 1% loss, a loss-reactive algorithm may collapse throughput to a fraction of the link capacity, while a model-based algorithm can push the send rate up without misjudging congestion. For large files like images and video, that difference shows up as "wait another 0.5 seconds" versus "wait 3 seconds".

Fine-Grained Tuning with Global Data

Cloudflare sees traffic from nearly every major network each day, including consumer ISPs, mobile networks, cloud providers, and low-Earth-orbit satellite links. It uses two sources of signal:

  • Passive logs: cover all traffic, recording send rates and acknowledgment timing, but only infer user experience.
  • Real User Measurement (RUM): browser-reported metrics such as page load time, closer to real experience but covering only a fraction of connections.

The approach uses a small amount of RUM data to calibrate, then extrapolates user-experience metrics from the full firehose of passive logs, enabling fine-grained experimentation and tuning. Using a Rust-based network stack that makes congestion control parameters easy to vary, Cloudflare has achieved about 10% faster responses versus baseline in some scenarios.

A Concrete Scenario: Image Downloads on a Weak Network

Suppose your homepage has an 800 KB hero image and the visitor is on a mobile network with mediocre 4G signal and elevated packet loss. With the baseline algorithm, loss triggers slowdown and repeated probing, so the image may take 2-3 seconds to appear; with tuned parameters, the link is filled earlier and more steadily, pulling load time under a second. For large images, video posters, and WebP galleries, a 10% throughput gain on top of first paint is very noticeable and directly improves the LCP metric in Core Web Vitals — see website performance optimization 2026.

Rollout and What It Means for Websites

The capability is running on free-tier QUIC traffic first and will expand to TCP and more complex customers through 2026. For site owners, any traffic served through a CDN benefits automatically, especially large files, images, and streaming.

If you are planning acceleration for your site, start with the Cloudflare setup guide, then refine with CDN cache strategy and multi-CDN load balancing. See the CDN Acceleration category for more.

How Site Owners Can Verify

These optimizations live inside the network stack, so users cannot toggle them directly, but two kinds of checks reveal the benefit:

  • Protocol level: confirm HTTP/3 (QUIC) is enabled, since the new parameters currently run on QUIC traffic first. Check with curl --http3 -I https://your-domain and look for alt-svc containing h3, or open the browser devtools network panel and look for the h3 protocol. HTTP/3 also saves one handshake round trip — see HTTP/3 and QUIC acceleration.
  • Real-user metrics: instrument LCP, FCP, and resource load times via RUM or Web Vitals, then compare before/after or across CDN edge regions. This beats single-shot synthetic tool scores.

Free-tier traffic is already covered for this QUIC portion; for entitlement boundaries see Cloudflare free vs paid plans.

Reference: QUIC specification RFC 9000 https://www.rfc-editor.org/rfc/rfc9000; BBR paper (Google Research) https://research.google/pubs/pub45646/; source https://blog.cloudflare.com/how-cloudflare-uses-the-worlds-greatest-collection-of-performance-data/