Roundhouse benchmark results

1. Throughput across targets

Each endpoint is its own chart. Bars are log-scaled; raw req/sec is shown at the right.

/articles

/articles/1

/articles/new

/articles.json

/articles/1.json

2. Lowerer dividend (ruby vs rails)

Same Ruby interpreter, same YJIT, same Puma — the only variable is whether the framework runtime is Rails or Roundhouse-emitted. The multiplier above each pair is the lift from the lowerer pipeline.

The same comparison on the JVM: emitted jruby beats stock rails-jruby by 17× on /articles, against 9× for ruby over rails on CRuby.

3. YJIT contribution

Each pair compares a cell with YJIT (ruby, rails) against its interpreter-only variant (ruby-int, rails-int). The bars show how far YJIT moves each.

4. Cost economics (req/sec per GB of RSS)

Throughput normalized by memory footprint — req/sec divided by RSS in GB. This reorders the raw throughput charts above: targets with small working sets rise and high-RSS targets fall. Bars are log-scaled. How much the metric matters depends on the deployment shape — most on metered or serverless surfaces, least on bare metal with memory headroom.

/articles

/articles/1

/articles/new

/articles.json

/articles/1.json

5. HTML → JSON multiplier

Per-target lift going from /articles (HTML) to /articles.json (JSON). The asymmetry isn't JSON-escape vs HTML-escape — both paths concatenate and escape strings. What JSON skips is the layout chain (csrf_meta_tag, importmap, asset-manifest scan, content_for slots) and the per-row view helpers (link_to, button_to, dom_id, tag.*) that the HTML index template invokes for every record. The chart below shows the multiplier per target.

rails-jruby's multiplier falls below 1× (0.8×): its JSON endpoint measures slower than its HTML one.

6. Memory footprint (RSS)

Max RSS observed across all endpoints, per target, sorted low to high. The field spans more than two orders of magnitude. The managed-heap cells (the JVM and .NET targets) run under a fixed heap budget rather than an organically-grown working set — see the note below the chart.

The JVM cells run under a fixed -Xmx512m heap, so the heap can't grab a host-dependent share of RAM; the rest is metaspace and code-cache, bounded by the app's class count rather than the box. Read the JVM bars as a pinned budget, not an organically-grown working set like the other cells.

The .NET cells are capped the same way — DOTNET_GCHeapHardLimit at 512 MB, the -Xmx analog — so their RSS is a budget the GC collects within, not the host-proportional reservation Server GC takes by default. NativeAOT (csharp-aot) carries no JIT or warmup, so its working set is the leanest of the managed-heap cells.

7. Latency (p50 / p99 at c=64)

Latencies are at c=64 concurrent connections; treat p50 as the median per-connection wait and p99 as the tail. Rows are ordered by measured median p50 across endpoints, fastest first.

All values in milliseconds. Lower is better.

Run-to-run stability

Each cell is timed 3 times and the charts above report the median run. This section reads per-run.json directly to show how far the individual runs strayed from it. Across all 85 cells the median run-to-run coefficient of variation in req/sec is 0.81% — the timed runs of a given cell agree to a fraction of a percent, so the reported medians aren't masking noise.

The 12 cells that vary by more than 3% are concentrated in the lowest-throughput cells, where a small absolute swing is a larger fraction of the rate; in 5 of them the first timed run is the slowest, consistent with residual warmup the fixed 20s warmup doesn't fully absorb:

req/sec per timed run; CV = standard deviation ÷ mean.

Generated 2026-09-10T22:36:37Z from summary.json.