ahmad hassan.
← All writing

November 20, 2026 · 13 min read

When Cellular Backhaul Goes Through Space

From in-flight Wi-Fi to direct-to-cell

If you’ve sent a message over Wi-Fi on a recent United flight, it may have traveled through a Starlink satellite. [1] Starlink also connects homes, businesses, and ships. As of September 30, 2026, SpaceX reported 13.6 million Starlink consumer subscribers. [2]

T-Mobile also uses Starlink to connect compatible phones directly to satellites in places without terrestrial coverage. The service supports texting, including text-to-911 where available. [3]

Could satellite backhaul become routine?

In late 2024 and early 2025, I started wondering whether Starlink could become a routine option for cellular backhaul. Your phone would connect to a terrestrial base station as usual, and a satellite link would carry that base station’s traffic to the mobile core. SpaceX had already reported that Starlink’s median U.S. peak-hour latency had fallen from about 49 ms to 33 ms. [4] I wanted to see how much more widely operators could use satellite backhaul as latency and packet loss came down and capacity increased.

Operators already used satellite backhaul in several settings. In 2021, KDDI selected Starlink to connect remote base stations in Japan. [5] Satellite links can also connect temporary cell sites at large events [6] or help restore cellular coverage after a disaster damages fiber. [7]

SpaceX’s move into mobile makes the question timely again. The FCC approved its EchoStar spectrum deal, and EchoStar still targets November 2027 for the final closing. [8] SpaceX has reportedly sought access to at least 80,000 U.S. cell sites. [9] Qualcomm will supply equipment and software for its planned terrestrial 5G network. [10]

Two wireless links in series

The cellular access link connects a phone to its serving base station. The Starlink backhaul carries traffic from the base station toward the mobile core through a satellite terminal at the site. The first link serves individual phones. The second carries their combined traffic.

Cellular access Starlink backhaul
Distance and propagation A 1 km radio path takes about 3 microseconds one way. At 550 km, the hop to a satellite directly overhead takes about 1.8 ms. The signal still has to reach a ground gateway. [5]
Measured one-way delay Under 1 ms at the physical layer in favorable conditions in our commercial millimeter-wave 5G measurements. [11] Baselines of 13 ms down and 17 ms up on a mostly idle path between a Starlink client and a cloud server. [12]
Scheduling and contention The base station assigns time and frequency resources to phones. Users in the same cell compete for those resources under load. [13] Starlink allocates bandwidth to the terminal based on demand. Traffic from the cell shares that allocation. [12]
Queueing and fairness Separate downlink queues for each phone’s radio bearers provide some isolation when the cell has capacity to spare. Under load, throughput shares depend on the scheduler. [14] One study found some delay isolation between flows and shared downlink loss. Another found unequal throughput for competing TCP flows. [12] [15]

Starlink backhaul. Even a stationary dish can change serving satellites. In SIGCOMM ’26 measurements, Starlink’s synchronized 15-second reconfigurations briefly interrupted packet delivery, reset bandwidth allocations, and could shift baseline delay. Changes also occurred when the serving satellite stayed the same. On a mostly idle link, UDP probe loss between reconfigurations averaged 0.02–0.03% and rose briefly at the boundary. [12] An earlier low-rate UDP study averaged 0.24% down and 1.24% up. [16] SatGuard measured 3–15% loss during short bursts in one web session. [17]

Cellular access. Per-user throughput varies with spectrum, radio conditions, and cell load. In our 5G measurements, handovers could temporarily disrupt application traffic. [18] Radio retransmissions help recover lost data and can add delay. [11] A campus Zoom study still observed more packet loss and jitter on cellular than on Wi-Fi in its dataset. [19]

What the sender sees. Either link can become the bottleneck. A backhaul slowdown can affect many users at once, even when their cell has radio capacity to spare. A sender sees the resulting delay and loss across the whole path. A jump in delay might mean a queue is growing on either link, or that Starlink’s baseline delay just changed.

Some questions we wanted to answer

  • What happens to packet delivery when a phone hands over to another cell just as Starlink reconfigures?
  • When does the bottleneck move between cellular access and satellite backhaul, and can we tell where packets are waiting?
  • How do changes in path delay and packet loss affect TCP throughput and queueing delay during large downloads?
  • How do those changes affect a video call, from the quality and smoothness of the video to the delay people experience?
  • How sensitive are cellular procedures such as establishing a data session to delay and loss on the backhaul?
  • Could the Starlink backhaul prioritize cellular control-plane messages and delay-sensitive traffic such as voice calls when capacity is limited?
  • Could the base station use backhaul capacity estimates to schedule users, and could the backhaul use the state of the radio queues to decide which packets to send first?
  • What changes to congestion control, queue management, or coordination between the two links could reduce the impact of handovers, reconfigurations, and packet loss?

Early insights

We tested two workloads with different demands. For a backlogged TCP transfer, such as a large download or software update, the sender always has more data ready and throughput largely determines how soon the transfer finishes. For real-time video, frames have to arrive in time for playback. Queueing delay during a bulk transfer can also matter to other traffic sharing the link. For example, a page loading over HTTP/2 can stall while the browser waits for a response it needs to request the next resource. [20]

For bulk transfers, we ran one TCP connection at a time, focusing on CUBIC and BBR, two widely deployed congestion controllers. [21] We measured both transfer throughput and queueing delay. For real-time video, we ran a WebRTC sender and receiver using Google Congestion Control (GCC) and recorded bitrate, frame delivery, and delay. [22] We used mmlight, a modified version of Mahimahi, to vary delay or inject loss while holding capacity at 24 Mbps in each direction. [23]

Prerequisite — TCP under packet loss

Starlink has lowered its latency, and longitudinal measurements have found periods of declining packet loss. The satellite hop still adds propagation delay compared with local Wi-Fi, cellular, or wired access. Reconfigurations can also cause brief loss spikes. [4] [27] [12]

CUBIC reduces its sending rate when it detects loss. [28] We ran CUBIC and BBR separately on a 24 Mbps link with a fixed 35 ms one-way delay. At 0.08% injected uplink loss, CUBIC used 32% of the link and BBR used 98%. At 0.32% loss, CUBIC used just 15%. BBR used more capacity at every nonzero loss rate in the sweep. Earlier experiments found that BBRv1 could also take a larger share from competing CUBIC flows at shallow queues. [26] We focused on BBR for the TCP delay experiments that follow.

CUBIC and BBR link utilization across nine injected uplink packet-loss settings.

BBR and Starlink reconfigurations

Starlink reconfigurations can shift the path’s baseline delay. [12] To see how BBR responds, we switched between 25 and 70 ms of one-way delay at different intervals, with zero injected loss. BBR used 79.5% of the link when delay changed every 12 seconds, compared with about 95% at 1- and 18-second intervals. Across the three 12-second runs, utilization was 77–81%. Utilization is achieved throughput divided by the configured 24 Mbps.

BBR uses its bandwidth and minimum round-trip time (RTT) estimates to decide how much data to keep in flight. After each of the first three delay increases in the trace, the minimum RTT estimate stayed near 51–52 ms for about nine seconds. An estimate that stays too low can leave too little data in flight to fill the link. Utilization fell during those periods and recovered after the estimate refreshed.

We used BBRv1 for these measurements. The current IETF draft specifies a 10-second minimum-RTT window. Google’s public BBR implementation retains that window and uses a separate five-second threshold for triggering an active RTT probe. With Starlink reconfigurations measured every 15 seconds, how those delay changes line up with BBR’s RTT sampling remains worth testing on newer builds. [24] [25]

BBR utilization across delay-change intervals, with the full 117-second run showing reconstructed one-way delay, smoothed TCP RTT, BBR minimum RTT, and utilization.

BBR and queueing delay

Many TCP applications depend on short response times. When you load a web page or an app makes a request to a server, time spent waiting in a queue can delay the next action. We wanted to see how changing the path delay affected queueing while a BBR flow kept the link busy.

We ran one backlogged TCP flow at a time on a 24 Mbps link with zero injected loss. We compared a fixed 35 ms one-way delay with a profile that alternated between 25 and 45 ms every 250 ms, with the same average configured delay. BBR sustained about 23.3 Mbps in both runs shown. The 95th-percentile uplink queueing delay rose from 54 ms with fixed delay to 97 ms in the second variable-delay run, the higher of the two repeats in that batch.

The 95th-percentile BBR queueing delay is 54 ms at a fixed 35 ms and 97 ms when delay alternates between 25 and 45 ms every 250 ms.

WebRTC and the timing of delay changes

A 20 ms swing in one-way delay was enough to push WebRTC below 1 Mbps at some intervals. To model end-to-end jitter on a path that crosses cellular access and satellite backhaul, we alternated between 25 and 45 ms on a 24 Mbps link with zero injected loss. Across one run per interval, WebRTC averaged 0.66 Mbps with changes every 250 ms, 0.43 Mbps with changes every second, and 12.5–18.6 Mbps at the other intervals. In a separate comparison, the sender started sending about 8 frames per second with 1-second changes and 48 at a fixed 45 ms.

GCC estimates whether a queue is growing by comparing the spacing between packet groups at the sender and receiver. A jump in the path’s baseline delay can stretch the gaps at the receiver and trigger a rate cut. [22] In the plotted 1-second run, the emulator’s uplink queue stayed near 1 ms. In a separate run with the same delay profile, the sender logged successful bandwidth probes and still sustained a low sending rate.

One possibility is that the fastest swings largely canceled out in GCC’s delay estimate, while longer intervals gave the sender time to recover. At 250 ms and 1 second, repeated increases may have triggered rate cuts before the sender could recover. We’d need GCC’s delay trend, overuse state, and target bitrate over time to test that explanation.

WebRTC sending rate across eight delay-change intervals with 25–45 ms one-way delay, with full sending-rate and emulator queueing-delay traces for 1- and 6-second intervals.

WebRTC under packet loss

We wanted to know how much of a call’s sending rate still goes to encoded video when packets are lost. WebRTC can send extra data as forward error correction (FEC), which lets the receiver recover some missing packets without waiting for retransmissions. We varied packet loss while holding the link at 24 Mbps and one-way delay at 45 ms. [29]

At 1.28% loss, the total RTP rate stayed near 17 Mbps while encoded video fell from 16.8 to 13.1 Mbps. The sender logged about 3.2 Mbps of FEC at that setting. We saw a similar pattern across 0.32–1.28% loss. At 2.56%, the total RTP rate fell to 11.5 Mbps, with 8.4 Mbps of encoded video.

Comparing zero loss with 1.28 percent injected loss, total RTP stays near 17 Mbps and encoded video falls from 16.8 to 13.1 Mbps.

What we learned and what remains

We wanted to understand what could go wrong when cellular access and satellite backhaul operate in series, then build something to keep applications responsive through handovers, reconfigurations, and packet loss.

The experiments answered some of those questions. BBR sustained more throughput than CUBIC under packet loss, and the timing of delay changes affected BBR’s link utilization and WebRTC’s sending rate. At the same average path delay, alternating delay increased queueing while TCP throughput stayed steady. With modest packet loss, WebRTC sent less encoded video as it allocated more bitrate to loss recovery. We still don’t know what happens when both links change at the same time, or how much prioritizing cellular signaling and delay-sensitive traffic would help.

In early srsRAN tests, some control-plane procedures took longer when we added delay to the emulated backhaul. We’d need more realistic experiments to confirm the latency impact and determine whether these procedures also fail more often.

We tried a TCP proxy that split a connection between the two links so each half could respond to its own link conditions. We tested it under fixed conditions and were still measuring the effects of delay and loss when I stepped away. Evaluating a design under the handovers and reconfigurations that motivated the project is still open.

My thesis defense, job search, and other projects took priority, and I don’t expect to pick this up again soon. If you’re interested in working on it, I’m happy to share the code, experiment setup, and notes.

Acknowledgments

Thanks to Ammar Tahir for his insights and for sharing his WebRTC setup, and to my advisor Feng Qian for his guidance.