How Cloudflare uses the world’s greatest collection of performance data to make the world’s fastest global network even faster

Congestion Control Gets a Data-Driven Upgrade

Cloudflare is rolling out a new approach to congestion control that leverages its visibility into global Internet traffic to squeeze more performance out of every connection. The company has announced that early results show average performance gains of 10% over its prior baseline, with plans to expand the improvements to all customers throughout 2026.

The work builds on Cloudflare’s ongoing overhaul of the software that accelerates its server fleet. But making individual servers faster is only half the battle — the network also has to handle massive, Internet-scale congestion events and route traffic to those servers efficiently. That’s where the new congestion control optimization comes in.

The Bottleneck Problem

When data travels across the Internet, it is broken into packets that hop from machine to machine over links with limited capacity. At any point along a path, one link will have the least available capacity, and that link becomes the bottleneck for the entire connection. The catch is that senders have no way of knowing which hop is the bottleneck, or even what route their packets will take. Each node in the network independently decides how to forward traffic, so packets between the same two endpoints can take entirely different paths.

Cloudflare has long used its own network visibility to route traffic between its datacenters efficiently via Argo Smart Routing, and to optimize connections to customer origins. But the last mile from a client device to Cloudflare’s nearest datacenter happens over third-party networks, where Cloudflare has no direct control.

The Congestion Control Arms Race

When too much data arrives at a node, packets either queue up — risking bufferbloat — or get dropped. TCP and QUIC respond to drops by retransmitting, which adds delay and can even worsen congestion. Early Internet history showed what happens when this is left unchecked: congestive collapse.

Congestion control algorithms were developed to prevent this, using different signals to estimate a safe sending rate. Traditional algorithms rely heavily on packet loss, detected when the sender never receives an acknowledgment (ACK) for a packet. Newer algorithms like BBR (Bottleneck Bandwidth and Round-trip propagation time), which Cloudflare has used extensively, build a model of connection capacity based on round-trip time estimates and loss information.

No single algorithm is optimal for every workload. Interactive traffic like video calls suffers when queues build up, while bulk downloads benefit from sending as much as possible. Different workloads, network conditions, and device types all call for different tuning — but only if you can see them.

Beyond Per-Connection Information

Existing congestion control algorithms are limited to signals gathered during a single connection’s lifetime. Cloudflare, however, observes traffic from essentially every major network on Earth, continuously. It knows the device type, network type (cellular, consumer ISP, cloud provider), and can see load patterns and identify systems that appear overloaded. It also knows which networks have stable properties, which suffer consistent cellular packet loss, and which use low-earth-orbit satellite links whose routes change every 15 seconds.

The company has migrated its networking stack to a new Rust-based platform that allows fine-grained experimentation with congestion control algorithm parameters. That gave it the flexibility to adapt algorithms per connection — but it needed data to know which parameters were working.

Measuring Real User Experience

To optimize for user experience, Cloudflare needs to measure it. The company gets two primary data sources. First, passive logs cover all traffic and show the rates at which data is sent and acknowledged. Second, Real User Measurement (RUM) data from supported web browsers records metrics like Page Load Time (PLT) — but RUM only covers a small fraction of connections.

The key insight: RUM data can be used to validate predictions built from passive logs alone. Cloudflare ran experiments on Free plan traffic, comparing two algorithms against a cubic baseline. The shape of the curves and, more importantly, the ratios between them were nearly identical when measured via RUM versus predicted from passive logs. That means a small RUM sample can validate findings, then the full passive log stream — covering vastly more connections — can be used for fine-grained network-wide optimization.

Next Steps and Rollout

Extrapolating from passive logs has limits, and Cloudflare is working with some of its largest customers to gain deeper visibility into network behavior from the client side. The result, Cloudflare says, is expected to be richer client experience insights for customers than any other platform can offer.

The experimental congestion control improvements are currently running on all Free tier QUIC traffic. As validation expands to more complex customer use cases and the work is extended to TCP traffic, Cloudflare plans to roll out the improvements to all customers across all traffic over the course of 2026 and beyond. The company is also running an early access program with select enterprises.