HTTP keep-alive throughput and not-modified latency for five BMLServerChannel builds, run side by side. Every number comes from the speed9 run report.
merge4e, the build now running on this server, ties merge4 and merge4d: about 127k requests/s over HTTP keep-alive with 32 connections (median of 9 rounds). It is not faster than merge4. The spread between merge4, merge4d and merge4e is small enough to be run-to-run noise.
Honest caveat: the not-modified latency differences (15.0–17.1 µs across builds) are within noise. Do not read them as a regression or an improvement. No HTTP round had errors.
| Build | HTTP keep-alive c=32, req/s (median) | Not-modified, median | Rounds with HTTP errors |
|---|---|---|---|
| merge3 | 121,289 | 15.2 µs | 0 |
| merge4 | 130,666 | 16.4 µs | 0 |
| merge4b | 122,764 | 16.1 µs | 0 |
| merge4d | 127,463 | 15.0 µs | 0 |
| merge4e | 127,101 | 17.1 µs | 0 |
Updated 2026-10-15 · bmlserverd merge4e · All numbers from the speed9 run report, none estimated