bench: splice() heuristics in the mix, 8K against the default 16K

Splicing after two 8K reads instead of 16K is never better in the mix
scenarios and costs 20% more CPU as an HTTP proxy on the 20 Mbit/s link:
the default stays.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Vladimir DubrovinandClaude Opus 5.5 committed 2026-10-14 20:19:49 +03:00
1 parent 21ba6e9d0c
commit cf3c491eec
3 files changed
+76

No files matched your search

+20
View File
@@ -1188,6 +1188,26 @@ Intel only: 10 Mbit/s, 10 ms and 0.5% loss each way, two CPUs for the proxy
the spread of the runs, as on the other links, and no setting moves the
latency.
### splice() heuristics in the mix
Intel only, `-s` (the default, `-s:16384,2,4096`) against `-s:8192`, which
splices after two reads of 8K rather than 16K: the mix scenarios, whose
answers have a median of 12K, unshaped and over the 20 Mbit/s, 5 ms, 0.5%
loss link with 5 client CPUs ([data/](data/) `skylake-mixsh-*.csv`); process
CPU per exchange, three runs each.
| scenario | -s | -s:8192 |
|---|---:|---:|
| mix-relay, unshaped | 62-65 µs | 61-65 µs |
| mix-http, unshaped | 96-100 µs | 93-104 µs |
| mix-connect, unshaped | 61-62 µs | 61-64 µs |
| mix-connect-3k, unshaped | 76-88 µs | 77-93 µs |
| mix-http-3k, 20 Mbit/s | 122-143 µs, p99 0.8 s | 152-172 µs, p99 3.1 s |
| mix-relay-3k, mix-connect-3k, 20 Mbit/s | 80-94 µs | 80-98 µs |
* Splicing the 8-16K exchanges too is never better and, as an HTTP proxy on
the slow link, costs 20% more CPU and a longer tail: the default stays.
## Earlier sessions
The morning session of 13 October ([data/](data/) `*-16k-*.csv`, `*-mix.csv`)