sjulia-coprime-pi-benchmark

name	sjulia-coprime-pi-benchmark
description	Use when asked to run or compare the coprime pi benchmark in the SubsetJuliaVM repo across Julia upstream, juliars, sjulia VM typed/untyped, and Python 3.14 via uv, especially when results must be reflected in docs/vm/BLOG.md.

Coprime π Five-Way Benchmark

Run the coprime π estimator in a 5-way comparison for N=5000 and N=10000, and reflect the results in docs/vm/BLOG.md.

Source files

Runtime	N=5000	N=10000
Julia upstream	benchmarks/calc_pi_n5000.jl	benchmarks/calc_pi_n10000.jl
sjulia VM typed	benchmarks/calc_pi_n5000_typed.jl	benchmarks/calc_pi_n10000_typed.jl
sjulia VM untyped	benchmarks/calc_pi_n5000.jl	benchmarks/calc_pi_n10000.jl
juliars	benchmarks/calc_pi_n5000_aot.jl	benchmarks/calc_pi_n10000_aot.jl
Python 3.14 (uv)	benchmarks/calc_pi_n5000.py	benchmarks/calc_pi_n10000.py

If the AoT files are missing, create them by keeping only calc_pi(N) at the end of the matching untyped source (no @time, no println).

Build prerequisites

cargo build --release -p subset_julia_vm --bin sjulia --features repl
cargo build --release -p subset_julia_vm --bin juliars --features aot

For warm VM runs comparable to the BLOG.md table, embed prelude/Base caches and rebuild:

./target/release/sjulia --precompile-prelude "$(pwd)/target/prelude_program_cache.bin"
./target/release/sjulia --precompile-base "$(pwd)/target/base_cache.bin"
SJULIA_PRELUDE_PROGRAM_CACHE="$(pwd)/target/prelude_program_cache.bin" \
SJULIA_BASE_CACHE="$(pwd)/target/base_cache.bin" \
  cargo build --release -p subset_julia_vm --bin sjulia --features repl

Commands

Run each benchmark at least twice and report the second (warm) wall time.

# 1. Julia upstream
time julia --startup-file=no benchmarks/calc_pi_n5000.jl
time julia --startup-file=no benchmarks/calc_pi_n10000.jl

# 2. sjulia VM typed
time ./target/release/sjulia benchmarks/calc_pi_n5000_typed.jl
time ./target/release/sjulia benchmarks/calc_pi_n10000_typed.jl

# 3. sjulia VM untyped
time ./target/release/sjulia benchmarks/calc_pi_n5000.jl
time ./target/release/sjulia benchmarks/calc_pi_n10000.jl

# 4. juliars
./target/release/juliars benchmarks/calc_pi_n5000_aot.jl \
  --minimal-prelude --emit-binary /tmp/calc_pi_n5000_aot
time /tmp/calc_pi_n5000_aot

./target/release/juliars benchmarks/calc_pi_n10000_aot.jl \
  --minimal-prelude --emit-binary /tmp/calc_pi_n10000_aot
time /tmp/calc_pi_n10000_aot

# 5. Python 3.14 (uv)
time uv run --python 3.14 --no-project benchmarks/calc_pi_n5000.py
time uv run --python 3.14 --no-project benchmarks/calc_pi_n10000.py

Reporting

Present results as a table:

Runtime	N=5000	N=10000
Julia upstream	...	...
sjulia VM typed	...	...
sjulia VM untyped	...	...
juliars	...	...
Python 3.14 (uv)	...	...

All runtimes should print the same π estimate:

• N=5000: 3.1413097643199746
• N=10000: 3.141534239016629

Reflecting results in docs/vm/BLOG.md

Update only the timing table inside the existing ### coprime π 推定 subsection. Do NOT replace the subsection wholesale: it also contains source listings (#### untyped / ... sub-subsections) that must be preserved, and the ## 実行パフォーマンス intro carries a runtime legend table shared by all benchmarks. The current BLOG table also has a Cython (uv) row — it is a hand-optimized 追試 (see ### Cython 版の追試); leave that row unchanged unless you re-measured Cython as well.

| Runtime | N=5000 | N=10000 |
|---|---:|---:|
| Julia upstream | X.XX s | X.XX s |
| sjulia VM typed | X.XX s | X.XX s |
| sjulia VM untyped | X.XX s | X.XX s |
| juliars | X.XX s | X.XX s |
| Python 3.14 (uv) | X.XX s | X.XX s |
| Cython (uv) | (据え置き・再計測時のみ更新) | (同左) |

Important caveats

• Warm runs only: The first sjulia VM run includes Base cache compilation and is much slower. Report the second real time.
• Cache embedding: BLOG.md numbers use a cache-embedded release binary. Without embedded caches, the CLI process time is dominated by Base compilation, not VM execution.
• AoT compile time is separate: juliars --emit-binary takes several seconds; report only the generated binary's execution time.
• No project overhead for Python: Use uv run --no-project so uv does not search for a pyproject.toml.
• Typed vs untyped gap: The typed variant uses explicit Int64/Float64 annotations. On simple integer loops, the untyped VM's runtime specialization can be competitive with or faster than the typed path.