CMSnap-LITE — public benchmark

Reproduce the published CMSnap-LITE benchmarks on your own machine. One script, nothing installed system-wide.

Published results: 32-core EPYC 9554 — 1M req/s, 10,000 sites, write scaling.

Requirements: Linux (x86_64 / arm64 / riscv64), curl, wrk (if wrk is missing, the script offers to install it via your package manager).

curl -fsSL https://raw.githubusercontent.com/zirill/cmsnap/main/bench.sh | sh

Or keep the script around:

curl -fLO https://raw.githubusercontent.com/zirill/cmsnap/main/bench.sh && chmod +x bench.sh
./bench.sh        # interactive menu
./bench.sh 2      # or run a test directly

The script downloads the official binary, unpacks the example site into ./run/ and drives wrk against it. It pins server and load to separate physical cores automatically (override with SERVER_CPUS / LOAD_CPUS).

Tests:

1. cached page, 1 site
2. cached page, 100 sites (multi-site Host routing)
3. page from database, 100 sites (cache off: SQLite read + render per request)
4. JSON API, cache off (/api/docs on every site)
5. write: public form POST into one site's database (honeypot + per-IP limiter on)
6. write across N sites — the POSTs fan out by Host, every site's own database takes inserts at once

Every knob is an environment variable:

SITES=1000 ./bench.sh 2        # a 1000-site node instead of the default 100
DURATION=120 CONNS=400 ./bench.sh 3

• SITES — how many sites the node runs in tests 2–4 (default 100); the node is rebuilt automatically when the number changes
• DURATION — seconds per measurement (default 60, after a 10 s warmup)
• CONNS — connections wrk keeps open (default 100)
• PORT — server port (default 8080)
• SERVER_CPUS / LOAD_CPUS — override the automatic core split

Results append to results.txt together with the machine details (binary version, kernel, CPU, RAM, pinning); rm -rf run results.txt removes every trace.

For publishable numbers: disable turbo, close background apps, run each test three times, keep the median.