Benchmark

monoruby is performance-focused, and performance is measured continuously rather than quoted from a README. This page covers the live dashboards, how to reproduce a measurement locally, and the historical one-off comparisons that used to live in the project wiki.

Live dashboards

Every push to master that touches the interpreter re-runs the benchmark suite and republishes the project portal:

The portal also hosts the two ruby/spec dashboards; those are covered in Compatibility.

Methodology, from .github/workflows/bench.yml:

• The suite is yjit-bench, run with --rss --harness=harness-warmup, monoruby against ruby --yjit on the same runner in the same job.
• The reference CRuby is 4.0.2 (ruby/setup-ruby), and monoruby is installed with cargo install --path monoruby --locked.
• Each benchmark gets a 400-second timeout per interpreter; a benchmark that times out or exits non-zero is recorded as a failure rather than silently dropped, and shows on the dashboard as a gap.
• The published ratio is × YJIT, higher = monoruby faster, 1× being parity. latest.json and data/history.csv next to each dashboard hold the raw numbers if you want to plot them yourself.

As a snapshot: on the x86-64 run of commit a13bfe0 (2026-08-31), 59 of the 76 benchmarks produced a ratio on both interpreters; over those the geometric mean was 1.28× YJIT, with monoruby ahead on 30 of them. The spread is wide in both directions — from ~16× on string_malloc_pressure down to ~0.3× on send_bmethod — which is the point of reading the dashboard rather than a single headline number.

Reproducing a measurement locally

The helper scripts in bin/ wrap the two harnesses the project uses. Most of them assume benchmark-driver and rbenv-managed reference Rubies; adjust the version strings inside to match what you have installed.

bin/compare is the one to reach for when you want to know whether a change you just made helped:

bin/compare                          # HEAD~1 vs HEAD, standard benchmarks
bin/compare abc1234 def5678          # two specific commits
bin/compare HEAD~5 HEAD app_fib.yml  # one benchmark

For a single script, a release build is enough:

cargo build --release
target/release/monoruby benchmark/app_fib.rb

Benchmark scripts and their benchmark-driver YAML configs live in benchmark/. Passing --no-jit gives you the VM-only baseline for the same script, which is often more informative than an absolute number.

Profiling

# Flame graph via Linux perf (needs ../FlameGraph)
bin/perf benchmark/app_fib.rb

# or by hand
cargo build --release --features perf
perf record target/release/monoruby benchmark/app_fib.rb
perf report

The perf feature makes monoruby emit perf-compatible symbol maps so JIT-compiled frames get names instead of raw addresses. .cargo/config.toml sets -Cforce-frame-pointers=yes globally, which is what makes the stacks walkable.

--features profile collects deopt and recompile statistics instead of a time profile — see Development and Build Options. Where optcarrot –opt spends its time is a worked example of both.

Historical measurements

The two comparisons below are one-off measurements previously published in the project wiki. They are kept for the record and are not re-measured; for current numbers use the live dashboards above.

optcarrot (April 2024)

Measured with optcarrot.

Rubies:

• ruby 3.4.0dev (2024-04-27T08:56:20Z master 9ea77cb351) [x86_64-linux]
• truffleruby 24.0.1, like ruby 3.2.2, Oracle GraalVM JVM [x86_64-linux]
• truffleruby 24.0.1, like ruby 3.2.2, Oracle GraalVM Native [x86_64-linux]
• monoruby 3e348afd4141c40978342e67ad26d42dc0b8d2a7

fps history, 0–3000 frames:

With --opt (optcarrot’s self-rewriting optimization mode):

yjit-bench (December 2024)

Speed ratio against truffleruby; higher is better. Measured with yjit-bench using --rss --harness=harness-warmup. Benchmark sources are from ruby/ruby’s benchmark/ and plb2.

Rubies:

• monoruby 0.3.0
• ruby 3.4.1 (2024-12-25 revision 48d4efcb85) +YJIT +PRISM [x86_64-linux]
• truffleruby 24.1.1, like ruby 3.2.4, Oracle GraalVM Native [x86_64-linux]

Raw data — execution time in milliseconds, resident set size (RSS) in MiB. monoruby/yjit and monoruby/truffle are time ratios; above 1 means monoruby is faster.

Machine

Both historical runs used the same machine:

• Architecture: x86_64
• CPU(s): 32 — 13th Gen Intel(R) Core(TM) i9-13900HX, 16 cores / 2 threads per core
• Caches (sum of all): L1d 768 KiB (16), L1i 512 KiB (16), L2 32 MiB (16), L3 36 MiB (1)