Summary

  • TurboTest predicts the throughput a full test would report, then decides whether the partial measurements are enough to stop.
  • On roughly one million NDT download tests, ε=15 saved 91.7% of transferred data and produced 18.6% median relative error.
  • The result is an estimate of a full test’s answer. It does not score an ISP’s advertised plan or a subscriber’s application experience.

Analysis

A speed test spends bandwidth to measure bandwidth. The faster the access link, the more bytes a fixed test can push through it. TURBOTEST asks the test to stop when its current evidence is good enough to predict what the full run would have reported.

Haarika Manda presented the work at LACNIC 46’s Internet Measurement Working Group on 21 October. The research paper appeared in the May 2026 USENIX NSDI proceedings. Its evaluation used about one million M-Lab NDT download tests recorded from April 2024 to March 2025; tests already cut short by the platform’s 250 MB cap were excluded.

The system separates two jobs that ordinary stopping rules blur. A first-stage regressor predicts full-run throughput from a partial test. A second-stage classifier decides whether to continue. It watches throughput alongside transport signals such as RTT, retransmissions, congestion-window size, bytes in flight and duplicate acknowledgements. A high-variability fallback lets a test finish when the early-stop decision is unreliable.

The parameter ε is the operator’s chosen error tolerance. At ε=15, the paper reports 91.7% data savings and 18.6% median relative error. The measured median overshoots the selected tolerance by 3.6 percentage points. At ε=20, the paper reports 95.1% savings with 24.6% median error. That is a trade-off curve, not a promise that every test saves the same amount or stays inside the same error bound.

The reference matters. The authors compare the early estimate with the throughput calculated from the full test. Their score answers whether TURBOTEST can approximate that completed measurement. It does not answer whether the completed test itself reflects the speed in a subscriber’s plan, the quality of a video call or a household’s experience. The paper explicitly leaves the question of what a speed test should report outside its scope.

Median error also hides the tail. The paper shows configurations that meet a 20% error bound at the median but fail to hold it at higher quantiles. A platform that publishes only the median can make a sharp-looking chart while the hardest tests carry a much larger error.

The missing evidence is a controlled or live rollout reporting error by speed tier and RTT, including high percentiles and fallback frequency. Those results would show whether the savings survive real deployment and changing network conditions.

Sources