Benchmarks

How Stowage performs under load. The full benchmark harness lives at benchmarks/. Numbers below come from the published results files at the time of v1.0.

#Setup

Captured 2026-04-26 to 2026-04-29.

• 16 concurrent workers per case.
• 15 s per case.
• Stowage and the upstream MinIO each constrained to 1 CPU and 200 MiB via cgroup limits in benchmarks/docker-compose.bench.yml.
• GOMEMLIMIT=180MiB on both processes so Go's GC stays inside the cgroup.
• Bench client and both servers shared the same host; loopback, no TLS.

#Dashboard endpoints

From benchmarks/results.md. Single-CPU, 16-concurrency:

Login is intentionally slow — argon2id m=65536 per hash, capped by a 10-attempts/15-min/IP limiter. Login concurrency cannot safely exceed 1 inside the 200 MiB container without OOM-killing the server.

#S3 proxy

From benchmarks/results-s3proxy.md:

#Proxy vs raw MinIO (head-to-head)

Under matched 1 CPU / 200 MiB constraints, the proxy adds +1–3 ms p50 / 0–11 % throughput for upstream-bound calls vs talking direct to MinIO. PutObject (both 1 KiB and 1 MiB) is faster than direct (+8 % and +18 % rps respectively). Synthesised paths (ListBuckets, scope reject, bad-sig reject) are much faster than MinIO's equivalent reject paths because the proxy answers without ever calling the upstream.

The detailed per-case comparison lives at benchmarks/results-comparison-proxy.md.

#Where the perf work landed

Three commits' worth, each driven by pprof under bench load:

#What you can do to make it faster

For the proxy:

1. Pool the response-stream copy buffer — alloc profiling shows io.copyBuffer at ~51 % of total bytes allocated on the read path. A sync.Pool of 32 KiB buffers would halve it.
2. Replace the outbound aws-sdk-go-v2/v4.Signer with a hand-rolled signer sharing the verifier's signing-key cache. ~3-4 % alloc win.
3. Audit DB on its own SQLite file. Today's audit and main writes share one mutex.

These are tracked in Roadmap.

#How to run the benchmarks yourself

cd benchmarks
docker compose -f docker-compose.bench.yml up -d --build
./run.sh

The harness is in benchmarks/bench.go (dashboard) and benchmarks/s3proxybench/ (S3 proxy). Output is JSON under benchmarks/results-*.json; the markdown summary files are generated by ./check.

#Calibration: what to read into these numbers

• Single 15 s sample per case. ±10–15 % run-to-run variance is normal.
• cgroup v1 doesn't enforce CPU as strictly as v2 — the "1 CPU" cap is approximate.
• Real deployments add TLS termination on both sides (~4 ms p50). The ratio between Stowage and direct shrinks as upstream RTT grows; the absolute Stowage overhead stays roughly constant.
• The four S3-shaped endpoints with a 1:1 MinIO mapping are the ones with a fair head-to-head comparison. Stowage's /api/me, /api/auth/config, etc. have no MinIO equivalent.