Sunday 20 September 2026, 03:05 PM
Decoupling TCP Prague from noisy RTTs to reduce L4S retransmissions by 20x
A July 2026 redesign of TCP Prague decouples segment sizing from noisy RTTs, relying on pacing rate to reduce L4S retransmissions by 20x on throttled links.
Let's talk about the exact moment a cloud gaming session falls apart. You are on a 5G connection, everything is buttery smooth, and then you step behind a concrete pillar. The signal fluctuates, the network panics, and your frame rate tanks. We usually just accept this as the reality of edge networks, but if you dig into the transport layer, the culprit for a lot of this stuttering is how our congestion control algorithms handle noisy data—specifically, Round Trip Time.
Apple and Comcast have been heavily backing the L4S architecture to get us to near-zero queuing delay, relying on TCP Prague to handle the traffic. It uses fine-grained ECN signals to keep packets flowing, which works well on a clean fiber line. But TCP Prague historically calculated its segment sizes using a mix of pacing rate and RTT. On a rural broadband connection or a cell edge, RTT bounces around. The algorithm reads that erratic data, guesses the wrong segment size, and dumps a burst of packets into the network. The router chokes, packets drop, and you end up caught in a loop of retransmissions.
A group of researchers from Nokia Bell Labs, the Max Planck Institute, Saarland University, and VU Amsterdam decided to just cut RTT out of the equation. They presented a paper last July at the ANRW 2026 workshop in Vienna called "Small Packets, Big Difference: L4S Congestion Control at Low Bandwidth." The idea was straightforward: stop trying to smooth out noisy RTT data and force the algorithm to rely strictly on the pacing rate.
They bypassed the usual ns-3 simulations and went straight for the Linux kernel. By engineering linear, exponential, and logarithmic mathematical schemes directly into a modified TCP Prague build, they tied segment sizing exclusively to pacing rate. Empirical tests on throttled links showed this dropped network retransmissions by a factor of 20.
Imagine the kind of interactive software we can build if we stop babysitting the transport layer. A 20x reduction in dropped packets opens up a lot of weird, highly scalable possibilities. If pacing rate is running the show and keeping the connection stable, we can push mixed reality applications over 5G RAN without worrying about the protocol tripping over a fluctuating signal. We can build multiplayer experiences that actually hold up on a train ride.
The telecom industry is already moving in this direction. 3GPP folded L4S into their 5G-Advanced Release 18 for exactly this kind of XR application, and CableLabs is baking it into Low Latency DOCSIS standards. Immunizing TCP Prague against noisy RTTs clears a practical hurdle for getting it into the mainline Linux kernel, mostly because ISPs can stop worrying about degraded links flooding their networks with retransmissions. By decoupling segment sizing from network noise, we finally have a transport layer that stays out of the way. Now we just get to figure out what to build on it.