Robert Zieliński Benchmarks Ruby Spinel vs JRuby vs CRuby (YJIT): 28 Ruby benchmarks 2026-11-13
Spinel , Matz’s ahead-of-time compiler for Ruby, ran 28 benchmark programs 20× faster than CRuby 4.0 with YJIT on the geometric mean, and faster on every one of them: from 4.4× (io_wordcount) to 334× (fib). It still cannot run a Rails app. This article has the numbers, the reason, and the pattern that lets a Rails app use Spinel anyway.
Spinel vs YJIT: 41× on computation, 15× on data structures and GC, 7.7× on the real-world programs. Against the plain interpreter: 16× overall.
JRuby 10.1 vs YJIT: 0.95× of YJIT’s speed after 15 warm-up runs, 0.14× when every run starts a new JVM.
Start-up: an empty program takes 0.77 ms as a Spinel binary, 5.3 ms on CRuby and 620 ms on JRuby.
Rails: Spinel refuses a string eval at compile time and never dispatches method_missing, so it cannot compile Rails. JRuby runs it. A Rails app can still call Spinel-compiled binaries for its pure, CPU-bound code.
Spinel parses a whole program with Prism, infers a static type for every value in it, emits C and hands that to the system C compiler. The result is a standalone native executable with no interpreter, no parser and no JIT inside, which needs only libc and libm. The price: everything has to be known when the binary is built.
Related measurements: Ruby vs JRuby on eight everyday fragments (long-running JRuby, memory, threads) and Ruby vs Rust on the same fragments (a compiled-language baseline).
Benchmark setup
Setup Details Programs the 28 programs in Spinel’s benchmark/ directory, unchanged, at upstream-equivalent commit 1c0cc5e35 Host 20-core ARM64 (aarch64) Linux server: 10 Cortex-X925 cores at 3.9 GHz and 10 Cortex-A725 at 2.8 GHz, performance governor, otherwise idle Pinning every runtime on the fast X925 cores: Spinel, CRuby and CRuby + YJIT on one core; JRuby on four (--cpuset-cpus=16-19), so its JIT and GC threads have cores of their own · containers pinned with docker --cpuset-cpus, the Spinel binary with taskset Repeats 3 trials per cell; the table and charts show the median. Trials are interleaved: all 28 programs for trial 1, then trial 2, then trial 3, so a disturbance lands in one trial, not all three. Median trial-to-trial spread: 3.4 %Spinel the native binary · a trial is the mean of 5 wall-clock runs CRuby 4.0.9, --disable-gems, with and without --yjit (RUBY_DESCRIPTION shows +YJIT) · a trial is the mean of 5 runs, timed inside the container JRuby 10.1.2 on OpenJDK 21, invokedynamic + JIT · cold : mean of 5 process runs, JVM start-up included · warm : the file loaded 20 times in one JVM, mean of the last 5, so JRuby after 15 warm-up runs, not a peak-throughput figure Correctness every program’s output on every runtime matches its .expected file in Spinel’s repo: 112 of 112 Empty program median of 3 trials of 5 starts: Spinel binary of nil 0.77 ms · ruby --disable-gems -e 0 5.3 ms · with --yjit 5.5 ms · jruby -e 0 620 ms Did not run none: all 28 ran on every runtime
Results: wall time per program
Spinel is the fastest runtime on every program except bigint_fib and pidigits, where warm JRuby is ahead. The three charts are the groups Spinel’s own README uses, each sorted by Spinel’s speedup over YJIT. The scale is logarithmic: one gridline is 10×.
Computation
Computation: wall time per benchmark in milliseconds, log scale, for Spinel, CRuby 4.0 with YJIT, the CRuby 4.0 interpreter and warm JRuby 10.1, sorted by Spinel's speedup over YJIT
Spinel: hover to isolate this series Spinel
CRuby 4.0 + YJIT: hover to isolate this series CRuby 4.0 + YJIT
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
vs YJIT
fib: Spinel 1.1 ms; CRuby + YJIT 354 ms; CRuby interpreter 345 ms; JRuby warm 1.84 s. Spinel is 334× faster than YJIT.
fib 334×
fib · Spinel: 1.1 ms (334× faster than YJIT) 1.1 ms
fib · CRuby + YJIT: 354 ms (the baseline) 354 ms
fib · CRuby interpreter: 345 ms (level with YJIT (within 5 %)) 345 ms
fib · JRuby warm: 1.84 s (5.2× slower than YJIT) 1.84 s
matmul: Spinel 3.2 ms; CRuby + YJIT 239 ms; CRuby interpreter 239 ms; JRuby warm 899 ms. Spinel is 75× faster than YJIT.
matmul 75×
matmul · Spinel: 3.2 ms (75× faster than YJIT) 3.2 ms
matmul · CRuby + YJIT: 239 ms (the baseline) 239 ms
matmul · CRuby interpreter: 239 ms (level with YJIT (within 5 %)) 239 ms
matmul · JRuby warm: 899 ms (3.8× slower than YJIT) 899 ms
tak: Spinel 6.5 ms; CRuby + YJIT 446 ms; CRuby interpreter 304 ms; JRuby warm 62.7 ms. Spinel is 68× faster than YJIT.
tak 68×
tak · Spinel: 6.5 ms (68× faster than YJIT) 6.5 ms
tak · CRuby + YJIT: 446 ms (the baseline) 446 ms
tak · CRuby interpreter: 304 ms (1.5× faster than YJIT) 304 ms
tak · JRuby warm: 62.7 ms (7.1× faster than YJIT) 62.7 ms
sieve: Spinel 7.2 ms; CRuby + YJIT 379 ms; CRuby interpreter 380 ms; JRuby warm 1.40 s. Spinel is 52× faster than YJIT.
sieve 52×
sieve · Spinel: 7.2 ms (52× faster than YJIT) 7.2 ms
sieve · CRuby + YJIT: 379 ms (the baseline) 379 ms
sieve · CRuby interpreter: 380 ms (level with YJIT (within 5 %)) 380 ms
sieve · JRuby warm: 1.40 s (3.7× slower than YJIT) 1.40 s
mandelbrot: Spinel 16.0 ms; CRuby + YJIT 774 ms; CRuby interpreter 774 ms; JRuby warm 2.93 s. Spinel is 48× faster than YJIT.
mandelbrot 48×
mandelbrot · Spinel: 16.0 ms (48× faster than YJIT) 16.0 ms
mandelbrot · CRuby + YJIT: 774 ms (the baseline) 774 ms
mandelbrot · CRuby interpreter: 774 ms (level with YJIT (within 5 %)) 774 ms
mandelbrot · JRuby warm: 2.93 s (3.8× slower than YJIT) 2.93 s
sudoku: Spinel 2.3 ms; CRuby + YJIT 109 ms; CRuby interpreter 71.1 ms; JRuby warm 577 ms. Spinel is 47× faster than YJIT.
sudoku 47×
sudoku · Spinel: 2.3 ms (47× faster than YJIT) 2.3 ms
sudoku · CRuby + YJIT: 109 ms (the baseline) 109 ms
sudoku · CRuby interpreter: 71.1 ms (1.5× faster than YJIT) 71.1 ms
sudoku · JRuby warm: 577 ms (5.3× slower than YJIT) 577 ms
tarai: Spinel 5.4 ms; CRuby + YJIT 245 ms; CRuby interpreter 237 ms; JRuby warm 69.1 ms. Spinel is 45× faster than YJIT.
tarai 45×
tarai · Spinel: 5.4 ms (45× faster than YJIT) 5.4 ms
tarai · CRuby + YJIT: 245 ms (the baseline) 245 ms
tarai · CRuby interpreter: 237 ms (level with YJIT (within 5 %)) 237 ms
tarai · JRuby warm: 69.1 ms (3.5× faster than YJIT) 69.1 ms
ackermann: Spinel 4.2 ms; CRuby + YJIT 176 ms; CRuby interpreter 305 ms; JRuby warm 112 ms. Spinel is 42× faster than YJIT.
ackermann 42×
ackermann · Spinel: 4.2 ms (42× faster than YJIT) 4.2 ms
ackermann · CRuby + YJIT: 176 ms (the baseline) 176 ms
ackermann · CRuby interpreter: 305 ms (1.7× slower than YJIT) 305 ms
ackermann · JRuby warm: 112 ms (1.6× faster than YJIT) 112 ms
partial_sums: Spinel 16.8 ms; CRuby + YJIT 695 ms; CRuby interpreter 693 ms; JRuby warm 2.39 s. Spinel is 41× faster than YJIT.
partial_sums 41×
partial_sums · Spinel: 16.8 ms (41× faster than YJIT) 16.8 ms
partial_sums · CRuby + YJIT: 695 ms (the baseline) 695 ms
partial_sums · CRuby interpreter: 693 ms (level with YJIT (within 5 %)) 693 ms
partial_sums · JRuby warm: 2.39 s (3.4× slower than YJIT) 2.39 s
life (Game of Life): Spinel 20.8 ms; CRuby + YJIT 830 ms; CRuby interpreter 480 ms; JRuby warm 430 ms. Spinel is 40× faster than YJIT.
life (Game of Life) 40×
life (Game of Life) · Spinel: 20.8 ms (40× faster than YJIT) 20.8 ms
life (Game of Life) · CRuby + YJIT: 830 ms (the baseline) 830 ms
life (Game of Life) · CRuby interpreter: 480 ms (1.7× faster than YJIT) 480 ms
life (Game of Life) · JRuby warm: 430 ms (1.9× faster than YJIT) 430 ms
nqueens: Spinel 4.9 ms; CRuby + YJIT 146 ms; CRuby interpreter 145 ms; JRuby warm 464 ms. Spinel is 30× faster than YJIT.
nqueens 30×
nqueens · Spinel: 4.9 ms (30× faster than YJIT) 4.9 ms
nqueens · CRuby + YJIT: 146 ms (the baseline) 146 ms
nqueens · CRuby interpreter: 145 ms (level with YJIT (within 5 %)) 145 ms
nqueens · JRuby warm: 464 ms (3.2× slower than YJIT) 464 ms
fannkuch: Spinel 1.1 ms; CRuby + YJIT 8.9 ms; CRuby interpreter 8.6 ms; JRuby warm 13.4 ms. Spinel is 8.0× faster than YJIT.
fannkuch 8.0×
fannkuch · Spinel: 1.1 ms (8.0× faster than YJIT) 1.1 ms
fannkuch · CRuby + YJIT: 8.9 ms (the baseline) 8.9 ms
fannkuch · CRuby interpreter: 8.6 ms (level with YJIT (within 5 %)) 8.6 ms
fannkuch · JRuby warm: 13.4 ms (1.5× slower than YJIT) 13.4 ms
fasta: Spinel 1.4 ms; CRuby + YJIT 9.3 ms; CRuby interpreter 8.5 ms; JRuby warm 28.2 ms. Spinel is 6.7× faster than YJIT.
fasta 6.7×
fasta · Spinel: 1.4 ms (6.7× faster than YJIT) 1.4 ms
fasta · CRuby + YJIT: 9.3 ms (the baseline) 9.3 ms
fasta · CRuby interpreter: 8.5 ms (1.1× faster than YJIT) 8.5 ms
fasta · JRuby warm: 28.2 ms (3.0× slower than YJIT) 28.2 ms
Computation: wall time, median of 3 trials, log scale (each gridline is 10×), sorted by Spinel's speedup over YJIT. JRuby warm is JRuby after 15 warm-up runs in one JVM. Hover a row or a legend entry to isolate it.
Computation: wall time per benchmark in milliseconds, log scale, for Spinel, CRuby 4.0 with YJIT, the CRuby 4.0 interpreter and warm JRuby 10.1, sorted by Spinel's speedup over YJIT
Spinel: hover to isolate this series Spinel
CRuby 4.0 + YJIT: hover to isolate this series CRuby 4.0 + YJIT
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
fib: Spinel 1.1 ms; CRuby + YJIT 354 ms; CRuby interpreter 345 ms; JRuby warm 1.84 s. Spinel is 334× faster than YJIT.
fib 334× vs YJIT
fib · Spinel: 1.1 ms (334× faster than YJIT) 1.1 ms
fib · CRuby + YJIT: 354 ms (the baseline) 354 ms
fib · CRuby interpreter: 345 ms (level with YJIT (within 5 %)) 345 ms
fib · JRuby warm: 1.84 s (5.2× slower than YJIT) 1.84 s
matmul: Spinel 3.2 ms; CRuby + YJIT 239 ms; CRuby interpreter 239 ms; JRuby warm 899 ms. Spinel is 75× faster than YJIT.
matmul 75× vs YJIT
matmul · Spinel: 3.2 ms (75× faster than YJIT) 3.2 ms
matmul · CRuby + YJIT: 239 ms (the baseline) 239 ms
matmul · CRuby interpreter: 239 ms (level with YJIT (within 5 %)) 239 ms
matmul · JRuby warm: 899 ms (3.8× slower than YJIT) 899 ms
tak: Spinel 6.5 ms; CRuby + YJIT 446 ms; CRuby interpreter 304 ms; JRuby warm 62.7 ms. Spinel is 68× faster than YJIT.
tak 68× vs YJIT
tak · Spinel: 6.5 ms (68× faster than YJIT) 6.5 ms
tak · CRuby + YJIT: 446 ms (the baseline) 446 ms
tak · CRuby interpreter: 304 ms (1.5× faster than YJIT) 304 ms
tak · JRuby warm: 62.7 ms (7.1× faster than YJIT) 62.7 ms
sieve: Spinel 7.2 ms; CRuby + YJIT 379 ms; CRuby interpreter 380 ms; JRuby warm 1.40 s. Spinel is 52× faster than YJIT.
sieve 52× vs YJIT
sieve · Spinel: 7.2 ms (52× faster than YJIT) 7.2 ms
sieve · CRuby + YJIT: 379 ms (the baseline) 379 ms
sieve · CRuby interpreter: 380 ms (level with YJIT (within 5 %)) 380 ms
sieve · JRuby warm: 1.40 s (3.7× slower than YJIT) 1.40 s
mandelbrot: Spinel 16.0 ms; CRuby + YJIT 774 ms; CRuby interpreter 774 ms; JRuby warm 2.93 s. Spinel is 48× faster than YJIT.
mandelbrot 48× vs YJIT
mandelbrot · Spinel: 16.0 ms (48× faster than YJIT) 16.0 ms
mandelbrot · CRuby + YJIT: 774 ms (the baseline) 774 ms
mandelbrot · CRuby interpreter: 774 ms (level with YJIT (within 5 %)) 774 ms
mandelbrot · JRuby warm: 2.93 s (3.8× slower than YJIT) 2.93 s
sudoku: Spinel 2.3 ms; CRuby + YJIT 109 ms; CRuby interpreter 71.1 ms; JRuby warm 577 ms. Spinel is 47× faster than YJIT.
sudoku 47× vs YJIT
sudoku · Spinel: 2.3 ms (47× faster than YJIT) 2.3 ms
sudoku · CRuby + YJIT: 109 ms (the baseline) 109 ms
sudoku · CRuby interpreter: 71.1 ms (1.5× faster than YJIT) 71.1 ms
sudoku · JRuby warm: 577 ms (5.3× slower than YJIT) 577 ms
tarai: Spinel 5.4 ms; CRuby + YJIT 245 ms; CRuby interpreter 237 ms; JRuby warm 69.1 ms. Spinel is 45× faster than YJIT.
tarai 45× vs YJIT
tarai · Spinel: 5.4 ms (45× faster than YJIT) 5.4 ms
tarai · CRuby + YJIT: 245 ms (the baseline) 245 ms
tarai · CRuby interpreter: 237 ms (level with YJIT (within 5 %)) 237 ms
tarai · JRuby warm: 69.1 ms (3.5× faster than YJIT) 69.1 ms
ackermann: Spinel 4.2 ms; CRuby + YJIT 176 ms; CRuby interpreter 305 ms; JRuby warm 112 ms. Spinel is 42× faster than YJIT.
ackermann 42× vs YJIT
ackermann · Spinel: 4.2 ms (42× faster than YJIT) 4.2 ms
ackermann · CRuby + YJIT: 176 ms (the baseline) 176 ms
ackermann · CRuby interpreter: 305 ms (1.7× slower than YJIT) 305 ms
ackermann · JRuby warm: 112 ms (1.6× faster than YJIT) 112 ms
partial_sums: Spinel 16.8 ms; CRuby + YJIT 695 ms; CRuby interpreter 693 ms; JRuby warm 2.39 s. Spinel is 41× faster than YJIT.
partial_sums 41× vs YJIT
partial_sums · Spinel: 16.8 ms (41× faster than YJIT) 16.8 ms
partial_sums · CRuby + YJIT: 695 ms (the baseline) 695 ms
partial_sums · CRuby interpreter: 693 ms (level with YJIT (within 5 %)) 693 ms
partial_sums · JRuby warm: 2.39 s (3.4× slower than YJIT) 2.39 s
life (Game of Life): Spinel 20.8 ms; CRuby + YJIT 830 ms; CRuby interpreter 480 ms; JRuby warm 430 ms. Spinel is 40× faster than YJIT.
life (Game of Life) 40× vs YJIT
life (Game of Life) · Spinel: 20.8 ms (40× faster than YJIT) 20.8 ms
life (Game of Life) · CRuby + YJIT: 830 ms (the baseline) 830 ms
life (Game of Life) · CRuby interpreter: 480 ms (1.7× faster than YJIT) 480 ms
life (Game of Life) · JRuby warm: 430 ms (1.9× faster than YJIT) 430 ms
nqueens: Spinel 4.9 ms; CRuby + YJIT 146 ms; CRuby interpreter 145 ms; JRuby warm 464 ms. Spinel is 30× faster than YJIT.
nqueens 30× vs YJIT
nqueens · Spinel: 4.9 ms (30× faster than YJIT) 4.9 ms
nqueens · CRuby + YJIT: 146 ms (the baseline) 146 ms
nqueens · CRuby interpreter: 145 ms (level with YJIT (within 5 %)) 145 ms
nqueens · JRuby warm: 464 ms (3.2× slower than YJIT) 464 ms
fannkuch: Spinel 1.1 ms; CRuby + YJIT 8.9 ms; CRuby interpreter 8.6 ms; JRuby warm 13.4 ms. Spinel is 8.0× faster than YJIT.
fannkuch 8.0× vs YJIT
fannkuch · Spinel: 1.1 ms (8.0× faster than YJIT) 1.1 ms
fannkuch · CRuby + YJIT: 8.9 ms (the baseline) 8.9 ms
fannkuch · CRuby interpreter: 8.6 ms (level with YJIT (within 5 %)) 8.6 ms
fannkuch · JRuby warm: 13.4 ms (1.5× slower than YJIT) 13.4 ms
fasta: Spinel 1.4 ms; CRuby + YJIT 9.3 ms; CRuby interpreter 8.5 ms; JRuby warm 28.2 ms. Spinel is 6.7× faster than YJIT.
fasta 6.7× vs YJIT
fasta · Spinel: 1.4 ms (6.7× faster than YJIT) 1.4 ms
fasta · CRuby + YJIT: 9.3 ms (the baseline) 9.3 ms
fasta · CRuby interpreter: 8.5 ms (1.1× faster than YJIT) 8.5 ms
fasta · JRuby warm: 28.2 ms (3.0× slower than YJIT) 28.2 ms
Computation: wall time, median of 3 trials, log scale (each gridline is 10×), sorted by Spinel's speedup over YJIT. JRuby warm is JRuby after 15 warm-up runs in one JVM.
Data structures and GC
Data structures & GC: wall time per benchmark in milliseconds, log scale, for Spinel, CRuby 4.0 with YJIT, the CRuby 4.0 interpreter and warm JRuby 10.1, sorted by Spinel's speedup over YJIT
Spinel: hover to isolate this series Spinel
CRuby 4.0 + YJIT: hover to isolate this series CRuby 4.0 + YJIT
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
vs YJIT
so_lists: Spinel 15.0 ms; CRuby + YJIT 389 ms; CRuby interpreter 228 ms; JRuby warm 248 ms. Spinel is 26× faster than YJIT.
so_lists 26×
so_lists · Spinel: 15.0 ms (26× faster than YJIT) 15.0 ms
so_lists · CRuby + YJIT: 389 ms (the baseline) 389 ms
so_lists · CRuby interpreter: 228 ms (1.7× faster than YJIT) 228 ms
so_lists · JRuby warm: 248 ms (1.6× faster than YJIT) 248 ms
rbtree: Spinel 18.0 ms; CRuby + YJIT 446 ms; CRuby interpreter 267 ms; JRuby warm 413 ms. Spinel is 25× faster than YJIT.
rbtree 25×
rbtree · Spinel: 18.0 ms (25× faster than YJIT) 18.0 ms
rbtree · CRuby + YJIT: 446 ms (the baseline) 446 ms
rbtree · CRuby interpreter: 267 ms (1.7× faster than YJIT) 267 ms
rbtree · JRuby warm: 413 ms (1.1× faster than YJIT) 413 ms
splay tree: Spinel 9.2 ms; CRuby + YJIT 170 ms; CRuby interpreter 94.4 ms; JRuby warm 315 ms. Spinel is 19× faster than YJIT.
splay tree 19×
splay tree · Spinel: 9.2 ms (19× faster than YJIT) 9.2 ms
splay tree · CRuby + YJIT: 170 ms (the baseline) 170 ms
splay tree · CRuby interpreter: 94.4 ms (1.8× faster than YJIT) 94.4 ms
splay tree · JRuby warm: 315 ms (1.9× slower than YJIT) 315 ms
gcbench: Spinel 182 ms; CRuby + YJIT 2.61 s; CRuby interpreter 1.78 s; JRuby warm 462 ms. Spinel is 14× faster than YJIT.
gcbench 14×
gcbench · Spinel: 182 ms (14× faster than YJIT) 182 ms
gcbench · CRuby + YJIT: 2.61 s (the baseline) 2.61 s
gcbench · CRuby interpreter: 1.78 s (1.5× faster than YJIT) 1.78 s
gcbench · JRuby warm: 462 ms (5.7× faster than YJIT) 462 ms
linked_list: Spinel 31.3 ms; CRuby + YJIT 336 ms; CRuby interpreter 177 ms; JRuby warm 331 ms. Spinel is 11× faster than YJIT.
linked_list 11×
linked_list · Spinel: 31.3 ms (11× faster than YJIT) 31.3 ms
linked_list · CRuby + YJIT: 336 ms (the baseline) 336 ms
linked_list · CRuby interpreter: 177 ms (1.9× faster than YJIT) 177 ms
linked_list · JRuby warm: 331 ms (level with YJIT (within 5 %)) 331 ms
binary_trees: Spinel 3.2 ms; CRuby + YJIT 32.2 ms; CRuby interpreter 21.2 ms; JRuby warm 50.8 ms. Spinel is 9.9× faster than YJIT.
binary_trees 9.9×
binary_trees · Spinel: 3.2 ms (9.9× faster than YJIT) 3.2 ms
binary_trees · CRuby + YJIT: 32.2 ms (the baseline) 32.2 ms
binary_trees · CRuby interpreter: 21.2 ms (1.5× faster than YJIT) 21.2 ms
binary_trees · JRuby warm: 50.8 ms (1.6× slower than YJIT) 50.8 ms
huffman: Spinel 4.4 ms; CRuby + YJIT 40.0 ms; CRuby interpreter 38.1 ms; JRuby warm 118 ms. Spinel is 9.1× faster than YJIT.
huffman 9.1×
huffman · Spinel: 4.4 ms (9.1× faster than YJIT) 4.4 ms
huffman · CRuby + YJIT: 40.0 ms (the baseline) 40.0 ms
huffman · CRuby interpreter: 38.1 ms (level with YJIT (within 5 %)) 38.1 ms
huffman · JRuby warm: 118 ms (2.9× slower than YJIT) 118 ms
Data structures & GC: wall time, median of 3 trials, log scale (each gridline is 10×), sorted by Spinel's speedup over YJIT. JRuby warm is JRuby after 15 warm-up runs in one JVM. Hover a row or a legend entry to isolate it.
Data structures & GC: wall time per benchmark in milliseconds, log scale, for Spinel, CRuby 4.0 with YJIT, the CRuby 4.0 interpreter and warm JRuby 10.1, sorted by Spinel's speedup over YJIT
Spinel: hover to isolate this series Spinel
CRuby 4.0 + YJIT: hover to isolate this series CRuby 4.0 + YJIT
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
so_lists: Spinel 15.0 ms; CRuby + YJIT 389 ms; CRuby interpreter 228 ms; JRuby warm 248 ms. Spinel is 26× faster than YJIT.
so_lists 26× vs YJIT
so_lists · Spinel: 15.0 ms (26× faster than YJIT) 15.0 ms
so_lists · CRuby + YJIT: 389 ms (the baseline) 389 ms
so_lists · CRuby interpreter: 228 ms (1.7× faster than YJIT) 228 ms
so_lists · JRuby warm: 248 ms (1.6× faster than YJIT) 248 ms
rbtree: Spinel 18.0 ms; CRuby + YJIT 446 ms; CRuby interpreter 267 ms; JRuby warm 413 ms. Spinel is 25× faster than YJIT.
rbtree 25× vs YJIT
rbtree · Spinel: 18.0 ms (25× faster than YJIT) 18.0 ms
rbtree · CRuby + YJIT: 446 ms (the baseline) 446 ms
rbtree · CRuby interpreter: 267 ms (1.7× faster than YJIT) 267 ms
rbtree · JRuby warm: 413 ms (1.1× faster than YJIT) 413 ms
splay tree: Spinel 9.2 ms; CRuby + YJIT 170 ms; CRuby interpreter 94.4 ms; JRuby warm 315 ms. Spinel is 19× faster than YJIT.
splay tree 19× vs YJIT
splay tree · Spinel: 9.2 ms (19× faster than YJIT) 9.2 ms
splay tree · CRuby + YJIT: 170 ms (the baseline) 170 ms
splay tree · CRuby interpreter: 94.4 ms (1.8× faster than YJIT) 94.4 ms
splay tree · JRuby warm: 315 ms (1.9× slower than YJIT) 315 ms
gcbench: Spinel 182 ms; CRuby + YJIT 2.61 s; CRuby interpreter 1.78 s; JRuby warm 462 ms. Spinel is 14× faster than YJIT.
gcbench 14× vs YJIT
gcbench · Spinel: 182 ms (14× faster than YJIT) 182 ms
gcbench · CRuby + YJIT: 2.61 s (the baseline) 2.61 s
gcbench · CRuby interpreter: 1.78 s (1.5× faster than YJIT) 1.78 s
gcbench · JRuby warm: 462 ms (5.7× faster than YJIT) 462 ms
linked_list: Spinel 31.3 ms; CRuby + YJIT 336 ms; CRuby interpreter 177 ms; JRuby warm 331 ms. Spinel is 11× faster than YJIT.
linked_list 11× vs YJIT
linked_list · Spinel: 31.3 ms (11× faster than YJIT) 31.3 ms
linked_list · CRuby + YJIT: 336 ms (the baseline) 336 ms
linked_list · CRuby interpreter: 177 ms (1.9× faster than YJIT) 177 ms
linked_list · JRuby warm: 331 ms (level with YJIT (within 5 %)) 331 ms
binary_trees: Spinel 3.2 ms; CRuby + YJIT 32.2 ms; CRuby interpreter 21.2 ms; JRuby warm 50.8 ms. Spinel is 9.9× faster than YJIT.
binary_trees 9.9× vs YJIT
binary_trees · Spinel: 3.2 ms (9.9× faster than YJIT) 3.2 ms
binary_trees · CRuby + YJIT: 32.2 ms (the baseline) 32.2 ms
binary_trees · CRuby interpreter: 21.2 ms (1.5× faster than YJIT) 21.2 ms
binary_trees · JRuby warm: 50.8 ms (1.6× slower than YJIT) 50.8 ms
huffman: Spinel 4.4 ms; CRuby + YJIT 40.0 ms; CRuby interpreter 38.1 ms; JRuby warm 118 ms. Spinel is 9.1× faster than YJIT.
huffman 9.1× vs YJIT
huffman · Spinel: 4.4 ms (9.1× faster than YJIT) 4.4 ms
huffman · CRuby + YJIT: 40.0 ms (the baseline) 40.0 ms
huffman · CRuby interpreter: 38.1 ms (level with YJIT (within 5 %)) 38.1 ms
huffman · JRuby warm: 118 ms (2.9× slower than YJIT) 118 ms
Data structures & GC: wall time, median of 3 trials, log scale (each gridline is 10×), sorted by Spinel's speedup over YJIT. JRuby warm is JRuby after 15 warm-up runs in one JVM.
Real-world programs
Real-world programs: wall time per benchmark in milliseconds, log scale, for Spinel, CRuby 4.0 with YJIT, the CRuby 4.0 interpreter and warm JRuby 10.1, sorted by Spinel's speedup over YJIT
Spinel: hover to isolate this series Spinel
CRuby 4.0 + YJIT: hover to isolate this series CRuby 4.0 + YJIT
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
vs YJIT
ao_render (ray tracer): Spinel 65.4 ms; CRuby + YJIT 2.03 s; CRuby interpreter 1.40 s; JRuby warm 2.13 s. Spinel is 31× faster than YJIT.
ao_render (ray tracer) 31×
ao_render (ray tracer) · Spinel: 65.4 ms (31× faster than YJIT) 65.4 ms
ao_render (ray tracer) · CRuby + YJIT: 2.03 s (the baseline) 2.03 s
ao_render (ray tracer) · CRuby interpreter: 1.40 s (1.4× faster than YJIT) 1.40 s
ao_render (ray tracer) · JRuby warm: 2.13 s (1.1× slower than YJIT) 2.13 s
json_parse: Spinel 22.8 ms; CRuby + YJIT 275 ms; CRuby interpreter 169 ms; JRuby warm 203 ms. Spinel is 12× faster than YJIT.
json_parse 12×
json_parse · Spinel: 22.8 ms (12× faster than YJIT) 22.8 ms
json_parse · CRuby + YJIT: 275 ms (the baseline) 275 ms
json_parse · CRuby interpreter: 169 ms (1.6× faster than YJIT) 169 ms
json_parse · JRuby warm: 203 ms (1.4× faster than YJIT) 203 ms
template engine: Spinel 60.3 ms; CRuby + YJIT 630 ms; CRuby interpreter 433 ms; JRuby warm 336 ms. Spinel is 10× faster than YJIT.
template engine 10×
template engine · Spinel: 60.3 ms (10× faster than YJIT) 60.3 ms
template engine · CRuby + YJIT: 630 ms (the baseline) 630 ms
template engine · CRuby interpreter: 433 ms (1.5× faster than YJIT) 433 ms
template engine · JRuby warm: 336 ms (1.9× faster than YJIT) 336 ms
bigint_fib: Spinel 1.1 ms; CRuby + YJIT 5.9 ms; CRuby interpreter 5.2 ms; JRuby warm 0.70 ms. Spinel is 5.4× faster than YJIT.
bigint_fib 5.4×
bigint_fib · Spinel: 1.1 ms (5.4× faster than YJIT) 1.1 ms
bigint_fib · CRuby + YJIT: 5.9 ms (the baseline) 5.9 ms
bigint_fib · CRuby interpreter: 5.2 ms (1.1× faster than YJIT) 5.2 ms
bigint_fib · JRuby warm: 0.70 ms (8.4× faster than YJIT) 0.70 ms
pidigits: Spinel 1.1 ms; CRuby + YJIT 5.9 ms; CRuby interpreter 5.8 ms; JRuby warm 0.60 ms. Spinel is 5.3× faster than YJIT.
pidigits 5.3×
pidigits · Spinel: 1.1 ms (5.3× faster than YJIT) 1.1 ms
pidigits · CRuby + YJIT: 5.9 ms (the baseline) 5.9 ms
pidigits · CRuby interpreter: 5.8 ms (level with YJIT (within 5 %)) 5.8 ms
pidigits · JRuby warm: 0.60 ms (9.8× faster than YJIT) 0.60 ms
csv_process: Spinel 98.3 ms; CRuby + YJIT 511 ms; CRuby interpreter 404 ms; JRuby warm 1.19 s. Spinel is 5.2× faster than YJIT.
csv_process 5.2×
csv_process · Spinel: 98.3 ms (5.2× faster than YJIT) 98.3 ms
csv_process · CRuby + YJIT: 511 ms (the baseline) 511 ms
csv_process · CRuby interpreter: 404 ms (1.3× faster than YJIT) 404 ms
csv_process · JRuby warm: 1.19 s (2.3× slower than YJIT) 1.19 s
str_concat: Spinel 1.1 ms; CRuby + YJIT 5.6 ms; CRuby interpreter 5.7 ms; JRuby warm 1.5 ms. Spinel is 5.0× faster than YJIT.
str_concat 5.0×
str_concat · Spinel: 1.1 ms (5.0× faster than YJIT) 1.1 ms
str_concat · CRuby + YJIT: 5.6 ms (the baseline) 5.6 ms
str_concat · CRuby interpreter: 5.7 ms (level with YJIT (within 5 %)) 5.7 ms
str_concat · JRuby warm: 1.5 ms (3.7× faster than YJIT) 1.5 ms
io_wordcount: Spinel 12.1 ms; CRuby + YJIT 53.4 ms; CRuby interpreter 41.8 ms; JRuby warm 76.2 ms. Spinel is 4.4× faster than YJIT.
io_wordcount 4.4×
io_wordcount · Spinel: 12.1 ms (4.4× faster than YJIT) 12.1 ms
io_wordcount · CRuby + YJIT: 53.4 ms (the baseline) 53.4 ms
io_wordcount · CRuby interpreter: 41.8 ms (1.3× faster than YJIT) 41.8 ms
io_wordcount · JRuby warm: 76.2 ms (1.4× slower than YJIT) 76.2 ms
Real-world programs: wall time, median of 3 trials, log scale (each gridline is 10×), sorted by Spinel's speedup over YJIT. JRuby warm is JRuby after 15 warm-up runs in one JVM. Hover a row or a legend entry to isolate it.
Real-world programs: wall time per benchmark in milliseconds, log scale, for Spinel, CRuby 4.0 with YJIT, the CRuby 4.0 interpreter and warm JRuby 10.1, sorted by Spinel's speedup over YJIT
Spinel: hover to isolate this series Spinel
CRuby 4.0 + YJIT: hover to isolate this series CRuby 4.0 + YJIT
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
ao_render (ray tracer): Spinel 65.4 ms; CRuby + YJIT 2.03 s; CRuby interpreter 1.40 s; JRuby warm 2.13 s. Spinel is 31× faster than YJIT.
ao_render (ray tracer) 31× vs YJIT
ao_render (ray tracer) · Spinel: 65.4 ms (31× faster than YJIT) 65.4 ms
ao_render (ray tracer) · CRuby + YJIT: 2.03 s (the baseline) 2.03 s
ao_render (ray tracer) · CRuby interpreter: 1.40 s (1.4× faster than YJIT) 1.40 s
ao_render (ray tracer) · JRuby warm: 2.13 s (1.1× slower than YJIT) 2.13 s
json_parse: Spinel 22.8 ms; CRuby + YJIT 275 ms; CRuby interpreter 169 ms; JRuby warm 203 ms. Spinel is 12× faster than YJIT.
json_parse 12× vs YJIT
json_parse · Spinel: 22.8 ms (12× faster than YJIT) 22.8 ms
json_parse · CRuby + YJIT: 275 ms (the baseline) 275 ms
json_parse · CRuby interpreter: 169 ms (1.6× faster than YJIT) 169 ms
json_parse · JRuby warm: 203 ms (1.4× faster than YJIT) 203 ms
template engine: Spinel 60.3 ms; CRuby + YJIT 630 ms; CRuby interpreter 433 ms; JRuby warm 336 ms. Spinel is 10× faster than YJIT.
template engine 10× vs YJIT
template engine · Spinel: 60.3 ms (10× faster than YJIT) 60.3 ms
template engine · CRuby + YJIT: 630 ms (the baseline) 630 ms
template engine · CRuby interpreter: 433 ms (1.5× faster than YJIT) 433 ms
template engine · JRuby warm: 336 ms (1.9× faster than YJIT) 336 ms
bigint_fib: Spinel 1.1 ms; CRuby + YJIT 5.9 ms; CRuby interpreter 5.2 ms; JRuby warm 0.70 ms. Spinel is 5.4× faster than YJIT.
bigint_fib 5.4× vs YJIT
bigint_fib · Spinel: 1.1 ms (5.4× faster than YJIT) 1.1 ms
bigint_fib · CRuby + YJIT: 5.9 ms (the baseline) 5.9 ms
bigint_fib · CRuby interpreter: 5.2 ms (1.1× faster than YJIT) 5.2 ms
bigint_fib · JRuby warm: 0.70 ms (8.4× faster than YJIT) 0.70 ms
pidigits: Spinel 1.1 ms; CRuby + YJIT 5.9 ms; CRuby interpreter 5.8 ms; JRuby warm 0.60 ms. Spinel is 5.3× faster than YJIT.
pidigits 5.3× vs YJIT
pidigits · Spinel: 1.1 ms (5.3× faster than YJIT) 1.1 ms
pidigits · CRuby + YJIT: 5.9 ms (the baseline) 5.9 ms
pidigits · CRuby interpreter: 5.8 ms (level with YJIT (within 5 %)) 5.8 ms
pidigits · JRuby warm: 0.60 ms (9.8× faster than YJIT) 0.60 ms
csv_process: Spinel 98.3 ms; CRuby + YJIT 511 ms; CRuby interpreter 404 ms; JRuby warm 1.19 s. Spinel is 5.2× faster than YJIT.
csv_process 5.2× vs YJIT
csv_process · Spinel: 98.3 ms (5.2× faster than YJIT) 98.3 ms
csv_process · CRuby + YJIT: 511 ms (the baseline) 511 ms
csv_process · CRuby interpreter: 404 ms (1.3× faster than YJIT) 404 ms
csv_process · JRuby warm: 1.19 s (2.3× slower than YJIT) 1.19 s
str_concat: Spinel 1.1 ms; CRuby + YJIT 5.6 ms; CRuby interpreter 5.7 ms; JRuby warm 1.5 ms. Spinel is 5.0× faster than YJIT.
str_concat 5.0× vs YJIT
str_concat · Spinel: 1.1 ms (5.0× faster than YJIT) 1.1 ms
str_concat · CRuby + YJIT: 5.6 ms (the baseline) 5.6 ms
str_concat · CRuby interpreter: 5.7 ms (level with YJIT (within 5 %)) 5.7 ms
str_concat · JRuby warm: 1.5 ms (3.7× faster than YJIT) 1.5 ms
io_wordcount: Spinel 12.1 ms; CRuby + YJIT 53.4 ms; CRuby interpreter 41.8 ms; JRuby warm 76.2 ms. Spinel is 4.4× faster than YJIT.
io_wordcount 4.4× vs YJIT
io_wordcount · Spinel: 12.1 ms (4.4× faster than YJIT) 12.1 ms
io_wordcount · CRuby + YJIT: 53.4 ms (the baseline) 53.4 ms
io_wordcount · CRuby interpreter: 41.8 ms (1.3× faster than YJIT) 41.8 ms
io_wordcount · JRuby warm: 76.2 ms (1.4× slower than YJIT) 76.2 ms
Real-world programs: wall time, median of 3 trials, log scale (each gridline is 10×), sorted by Spinel's speedup over YJIT. JRuby warm is JRuby after 15 warm-up runs in one JVM.
How to read the Spinel numbers
fib (334×) is not a per-call number. Its whole call tree is a pure function of a literal, so the C compiler folds most of it at build time. It shows what a compiler can do when it sees the whole program, not how fast one Ruby method call is.
Short cells include process start-up. Every time includes starting the process: 0.77 ms for an empty Spinel binary, 5.5 ms for an empty ruby --yjit. That start-up is more than half of Spinel’s time on 6 programs (fib 73%, bigint_fib 70%, fannkuch 69%, pidigits 69%, str_concat 68% and fasta 56%). On bigint_fib, pidigits and str_concat, YJIT’s time is within 1 ms of its own empty start, so those ratios say almost nothing about the code.
YJIT rarely helps these short programs on this host. It is more than 5 % faster than the interpreter on 1 of 28 (best: ackermann, 1.7× the interpreter’s speed), more than 5 % slower on 16 (worst: linked_list, 0.52×) and within 5 % on the other 11. Spinel’s README reports a 5–8× YJIT gain on its call-heavy programs (ackermann, tarai, tak, rbtree) on its own machine; that gain does not show here, so every multiple in this article is against YJIT as measured on this ARM64 host.
The smallest gains: io_wordcount (4.4×), str_concat (5.0×) and csv_process (5.2×). io_wordcount and csv_process spend their time splitting text into lines and fields, work CRuby already does in C. str_concat is mostly process start-up on both sides (see above).
Geometric mean: Spinel, CRuby and JRuby against YJIT
Geometric-mean speed relative to CRuby 4.0 with YJIT, log scale, per benchmark group and over all programs: Spinel, the CRuby interpreter, warm JRuby and cold JRuby
Spinel: hover to isolate this series Spinel
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
JRuby 10.1 cold: hover to isolate this series JRuby 10.1 cold
0.01×
0.1×
1×
10×
100×
◂ slower faster than YJIT ▸
Computation (13 programs), geometric mean against YJIT = 1×: Spinel 41×; CRuby interpreter 1.1×; JRuby warm 0.59×; JRuby cold 0.16×.
Computation 13 programs
Computation · Spinel: 41× as fast as YJIT (geometric mean of 13 programs) 41×
Computation · CRuby interpreter: 1.1× as fast as YJIT (geometric mean of 13 programs) 1.1×
Computation · JRuby warm: 0.59× of YJIT's speed (geometric mean of 13 programs) 0.59×
Computation · JRuby cold: 0.16× of YJIT's speed (geometric mean of 13 programs) 0.16×
Data structures & GC (7 programs), geometric mean against YJIT = 1×: Spinel 15×; CRuby interpreter 1.6×; JRuby warm 1.0×; JRuby cold 0.23×.
Data structures & GC 7 programs
Data structures & GC · Spinel: 15× as fast as YJIT (geometric mean of 7 programs) 15×
Data structures & GC · CRuby interpreter: 1.6× as fast as YJIT (geometric mean of 7 programs) 1.6×
Data structures & GC · JRuby warm: level with YJIT (within 5 %) (geometric mean of 7 programs) 1.0×
Data structures & GC · JRuby cold: 0.23× of YJIT's speed (geometric mean of 7 programs) 0.23×
Real-world programs (8 programs), geometric mean against YJIT = 1×: Spinel 7.7×; CRuby interpreter 1.3×; JRuby warm 2.0×; JRuby cold 0.08×.
Real-world programs 8 programs
Real-world programs · Spinel: 7.7× as fast as YJIT (geometric mean of 8 programs) 7.7×
Real-world programs · CRuby interpreter: 1.3× as fast as YJIT (geometric mean of 8 programs) 1.3×
Real-world programs · JRuby warm: 2.0× as fast as YJIT (geometric mean of 8 programs) 2.0×
Real-world programs · JRuby cold: 0.08× of YJIT's speed (geometric mean of 8 programs) 0.08×
All programs (28 programs), geometric mean against YJIT = 1×: Spinel 20×; CRuby interpreter 1.2×; JRuby warm 0.95×; JRuby cold 0.14×.
All programs 28 programs
All programs · Spinel: 20× as fast as YJIT (geometric mean of 28 programs) 20×
All programs · CRuby interpreter: 1.2× as fast as YJIT (geometric mean of 28 programs) 1.2×
All programs · JRuby warm: level with YJIT (within 5 %) (geometric mean of 28 programs) 0.95×
All programs · JRuby cold: 0.14× of YJIT's speed (geometric mean of 28 programs) 0.14×
Geometric mean of the per-program speedup against CRuby 4.0 + YJIT, log scale. Right of the 1× line is faster than YJIT, left is slower. Hover a row or a legend entry to isolate it.
Geometric-mean speed relative to CRuby 4.0 with YJIT, log scale, per benchmark group and over all programs: Spinel, the CRuby interpreter, warm JRuby and cold JRuby
Spinel: hover to isolate this series Spinel
CRuby 4.0 interpreter: hover to isolate this series CRuby 4.0 interpreter
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
JRuby 10.1 cold: hover to isolate this series JRuby 10.1 cold
0.01×
0.1×
1×
10×
100×
◂ slower faster than YJIT ▸
Computation (13 programs), geometric mean against YJIT = 1×: Spinel 41×; CRuby interpreter 1.1×; JRuby warm 0.59×; JRuby cold 0.16×.
Computation · 13 programs
Computation · Spinel: 41× as fast as YJIT (geometric mean of 13 programs) 41×
Computation · CRuby interpreter: 1.1× as fast as YJIT (geometric mean of 13 programs) 1.1×
Computation · JRuby warm: 0.59× of YJIT's speed (geometric mean of 13 programs) 0.59×
Computation · JRuby cold: 0.16× of YJIT's speed (geometric mean of 13 programs) 0.16×
Data structures & GC (7 programs), geometric mean against YJIT = 1×: Spinel 15×; CRuby interpreter 1.6×; JRuby warm 1.0×; JRuby cold 0.23×.
Data structures & GC · 7 programs
Data structures & GC · Spinel: 15× as fast as YJIT (geometric mean of 7 programs) 15×
Data structures & GC · CRuby interpreter: 1.6× as fast as YJIT (geometric mean of 7 programs) 1.6×
Data structures & GC · JRuby warm: level with YJIT (within 5 %) (geometric mean of 7 programs) 1.0×
Data structures & GC · JRuby cold: 0.23× of YJIT's speed (geometric mean of 7 programs) 0.23×
Real-world programs (8 programs), geometric mean against YJIT = 1×: Spinel 7.7×; CRuby interpreter 1.3×; JRuby warm 2.0×; JRuby cold 0.08×.
Real-world programs · 8 programs
Real-world programs · Spinel: 7.7× as fast as YJIT (geometric mean of 8 programs) 7.7×
Real-world programs · CRuby interpreter: 1.3× as fast as YJIT (geometric mean of 8 programs) 1.3×
Real-world programs · JRuby warm: 2.0× as fast as YJIT (geometric mean of 8 programs) 2.0×
Real-world programs · JRuby cold: 0.08× of YJIT's speed (geometric mean of 8 programs) 0.08×
All programs (28 programs), geometric mean against YJIT = 1×: Spinel 20×; CRuby interpreter 1.2×; JRuby warm 0.95×; JRuby cold 0.14×.
All programs · 28 programs
All programs · Spinel: 20× as fast as YJIT (geometric mean of 28 programs) 20×
All programs · CRuby interpreter: 1.2× as fast as YJIT (geometric mean of 28 programs) 1.2×
All programs · JRuby warm: level with YJIT (within 5 %) (geometric mean of 28 programs) 0.95×
All programs · JRuby cold: 0.14× of YJIT's speed (geometric mean of 28 programs) 0.14×
Geometric mean of the per-program speedup against CRuby 4.0 + YJIT, log scale. Right of the 1× line is faster than YJIT, left is slower.
With YJIT as 1×, Spinel is at 20× over all 28 programs. On this host the plain interpreter is 1.2× as fast as YJIT on the geometric mean.
JRuby: start-up vs steady state
A cold JRuby run never finishes under 622 ms, because an empty program on a new JVM already takes 620 ms. That start-up is at least half of the cold time on 19 of the 28 programs; cold runs span 622 ms–3.04 s. A Spinel binary starts in 0.77 ms, 805× faster, and ruby -e 0 in 5.3 ms.
Per benchmark, log scale, sorted by warm JRuby time: Spinel, JRuby warm (same JVM, mean of the last 5 of 20 runs) and JRuby cold (a new JVM per run). The bar between warm and cold is JVM start-up plus warm-up.
Spinel: hover to isolate this series Spinel
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
JRuby 10.1 cold: hover to isolate this series JRuby 10.1 cold
between warm and cold: start-up + warm-up
an empty program on a new JVM: 620 ms
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
pidigits: Spinel 1.1 ms; JRuby warm 0.60 ms; JRuby cold 644 ms. A cold JRuby run takes 580× the Spinel time.
pidigits
pidigits · JRuby cold: 644 ms
pidigits · JRuby warm: 0.60 ms
pidigits · Spinel: 1.1 ms
Spinel 1.1 ms · JRuby warm 0.60 ms · JRuby cold 644 ms · cold = 580× Spinel
bigint_fib: Spinel 1.1 ms; JRuby warm 0.70 ms; JRuby cold 622 ms. A cold JRuby run takes 565× the Spinel time.
bigint_fib
bigint_fib · JRuby cold: 622 ms
bigint_fib · JRuby warm: 0.70 ms
bigint_fib · Spinel: 1.1 ms
Spinel 1.1 ms · JRuby warm 0.70 ms · JRuby cold 622 ms · cold = 565× Spinel
str_concat: Spinel 1.1 ms; JRuby warm 1.5 ms; JRuby cold 636 ms. A cold JRuby run takes 563× the Spinel time.
str_concat
str_concat · JRuby cold: 636 ms
str_concat · JRuby warm: 1.5 ms
str_concat · Spinel: 1.1 ms
Spinel 1.1 ms · JRuby warm 1.5 ms · JRuby cold 636 ms · cold = 563× Spinel
fannkuch: Spinel 1.1 ms; JRuby warm 13.4 ms; JRuby cold 706 ms. A cold JRuby run takes 636× the Spinel time.
fannkuch
fannkuch · JRuby cold: 706 ms
fannkuch · JRuby warm: 13.4 ms
fannkuch · Spinel: 1.1 ms
Spinel 1.1 ms · JRuby warm 13.4 ms · JRuby cold 706 ms · cold = 636× Spinel
fasta: Spinel 1.4 ms; JRuby warm 28.2 ms; JRuby cold 742 ms. A cold JRuby run takes 537× the Spinel time.
fasta
fasta · JRuby cold: 742 ms
fasta · JRuby warm: 28.2 ms
fasta · Spinel: 1.4 ms
Spinel 1.4 ms · JRuby warm 28.2 ms · JRuby cold 742 ms · cold = 537× Spinel
binary_trees: Spinel 3.2 ms; JRuby warm 50.8 ms; JRuby cold 714 ms. A cold JRuby run takes 220× the Spinel time.
binary_trees
binary_trees · JRuby cold: 714 ms
binary_trees · JRuby warm: 50.8 ms
binary_trees · Spinel: 3.2 ms
Spinel 3.2 ms · JRuby warm 50.8 ms · JRuby cold 714 ms · cold = 220× Spinel
tak: Spinel 6.5 ms; JRuby warm 62.7 ms; JRuby cold 702 ms. A cold JRuby run takes 107× the Spinel time.
tak
tak · JRuby cold: 702 ms
tak · JRuby warm: 62.7 ms
tak · Spinel: 6.5 ms
Spinel 6.5 ms · JRuby warm 62.7 ms · JRuby cold 702 ms · cold = 107× Spinel
tarai: Spinel 5.4 ms; JRuby warm 69.1 ms; JRuby cold 689 ms. A cold JRuby run takes 127× the Spinel time.
tarai
tarai · JRuby cold: 689 ms
tarai · JRuby warm: 69.1 ms
tarai · Spinel: 5.4 ms
Spinel 5.4 ms · JRuby warm 69.1 ms · JRuby cold 689 ms · cold = 127× Spinel
io_wordcount: Spinel 12.1 ms; JRuby warm 76.2 ms; JRuby cold 786 ms. A cold JRuby run takes 65× the Spinel time.
io_wordcount
io_wordcount · JRuby cold: 786 ms
io_wordcount · JRuby warm: 76.2 ms
io_wordcount · Spinel: 12.1 ms
Spinel 12.1 ms · JRuby warm 76.2 ms · JRuby cold 786 ms · cold = 65× Spinel
ackermann: Spinel 4.2 ms; JRuby warm 112 ms; JRuby cold 783 ms. A cold JRuby run takes 186× the Spinel time.
ackermann
ackermann · JRuby cold: 783 ms
ackermann · JRuby warm: 112 ms
ackermann · Spinel: 4.2 ms
Spinel 4.2 ms · JRuby warm 112 ms · JRuby cold 783 ms · cold = 186× Spinel
huffman: Spinel 4.4 ms; JRuby warm 118 ms; JRuby cold 1.08 s. A cold JRuby run takes 245× the Spinel time.
huffman
huffman · JRuby cold: 1.08 s
huffman · JRuby warm: 118 ms
huffman · Spinel: 4.4 ms
Spinel 4.4 ms · JRuby warm 118 ms · JRuby cold 1.08 s · cold = 245× Spinel
json_parse: Spinel 22.8 ms; JRuby warm 203 ms; JRuby cold 888 ms. A cold JRuby run takes 39× the Spinel time.
json_parse
json_parse · JRuby cold: 888 ms
json_parse · JRuby warm: 203 ms
json_parse · Spinel: 22.8 ms
Spinel 22.8 ms · JRuby warm 203 ms · JRuby cold 888 ms · cold = 39× Spinel
so_lists: Spinel 15.0 ms; JRuby warm 248 ms; JRuby cold 858 ms. A cold JRuby run takes 57× the Spinel time.
so_lists
so_lists · JRuby cold: 858 ms
so_lists · JRuby warm: 248 ms
so_lists · Spinel: 15.0 ms
Spinel 15.0 ms · JRuby warm 248 ms · JRuby cold 858 ms · cold = 57× Spinel
splay tree: Spinel 9.2 ms; JRuby warm 315 ms; JRuby cold 1.12 s. A cold JRuby run takes 123× the Spinel time.
splay tree
splay tree · JRuby cold: 1.12 s
splay tree · JRuby warm: 315 ms
splay tree · Spinel: 9.2 ms
Spinel 9.2 ms · JRuby warm 315 ms · JRuby cold 1.12 s · cold = 123× Spinel
linked_list: Spinel 31.3 ms; JRuby warm 331 ms; JRuby cold 992 ms. A cold JRuby run takes 32× the Spinel time.
linked_list
linked_list · JRuby cold: 992 ms
linked_list · JRuby warm: 331 ms
linked_list · Spinel: 31.3 ms
Spinel 31.3 ms · JRuby warm 331 ms · JRuby cold 992 ms · cold = 32× Spinel
template engine: Spinel 60.3 ms; JRuby warm 336 ms; JRuby cold 1.07 s. A cold JRuby run takes 18× the Spinel time.
template engine
template engine · JRuby cold: 1.07 s
template engine · JRuby warm: 336 ms
template engine · Spinel: 60.3 ms
Spinel 60.3 ms · JRuby warm 336 ms · JRuby cold 1.07 s · cold = 18× Spinel
rbtree: Spinel 18.0 ms; JRuby warm 413 ms; JRuby cold 1.03 s. A cold JRuby run takes 57× the Spinel time.
rbtree
rbtree · JRuby cold: 1.03 s
rbtree · JRuby warm: 413 ms
rbtree · Spinel: 18.0 ms
Spinel 18.0 ms · JRuby warm 413 ms · JRuby cold 1.03 s · cold = 57× Spinel
life (Game of Life): Spinel 20.8 ms; JRuby warm 430 ms; JRuby cold 1.27 s. A cold JRuby run takes 61× the Spinel time.
life (Game of Life)
life (Game of Life) · JRuby cold: 1.27 s
life (Game of Life) · JRuby warm: 430 ms
life (Game of Life) · Spinel: 20.8 ms
Spinel 20.8 ms · JRuby warm 430 ms · JRuby cold 1.27 s · cold = 61× Spinel
gcbench: Spinel 182 ms; JRuby warm 462 ms; JRuby cold 1.44 s. A cold JRuby run takes 7.9× the Spinel time.
gcbench
gcbench · JRuby cold: 1.44 s
gcbench · JRuby warm: 462 ms
gcbench · Spinel: 182 ms
Spinel 182 ms · JRuby warm 462 ms · JRuby cold 1.44 s · cold = 7.9× Spinel
nqueens: Spinel 4.9 ms; JRuby warm 464 ms; JRuby cold 1.41 s. A cold JRuby run takes 288× the Spinel time.
nqueens
nqueens · JRuby cold: 1.41 s
nqueens · JRuby warm: 464 ms
nqueens · Spinel: 4.9 ms
Spinel 4.9 ms · JRuby warm 464 ms · JRuby cold 1.41 s · cold = 288× Spinel
sudoku: Spinel 2.3 ms; JRuby warm 577 ms; JRuby cold 1.06 s. A cold JRuby run takes 458× the Spinel time.
sudoku
sudoku · JRuby cold: 1.06 s
sudoku · JRuby warm: 577 ms
sudoku · Spinel: 2.3 ms
Spinel 2.3 ms · JRuby warm 577 ms · JRuby cold 1.06 s · cold = 458× Spinel
matmul: Spinel 3.2 ms; JRuby warm 899 ms; JRuby cold 1.24 s. A cold JRuby run takes 389× the Spinel time.
matmul
matmul · JRuby cold: 1.24 s
matmul · JRuby warm: 899 ms
matmul · Spinel: 3.2 ms
Spinel 3.2 ms · JRuby warm 899 ms · JRuby cold 1.24 s · cold = 389× Spinel
csv_process: Spinel 98.3 ms; JRuby warm 1.19 s; JRuby cold 2.20 s. A cold JRuby run takes 22× the Spinel time.
csv_process
csv_process · JRuby cold: 2.20 s
csv_process · JRuby warm: 1.19 s
csv_process · Spinel: 98.3 ms
Spinel 98.3 ms · JRuby warm 1.19 s · JRuby cold 2.20 s · cold = 22× Spinel
sieve: Spinel 7.2 ms; JRuby warm 1.40 s; JRuby cold 1.27 s. A cold JRuby run takes 176× the Spinel time.
sieve
sieve · JRuby cold: 1.27 s
sieve · JRuby warm: 1.40 s
sieve · Spinel: 7.2 ms
Spinel 7.2 ms · JRuby warm 1.40 s · JRuby cold 1.27 s · cold = 176× Spinel
fib: Spinel 1.1 ms; JRuby warm 1.84 s; JRuby cold 2.27 s. A cold JRuby run takes 2144× the Spinel time.
fib
fib · JRuby cold: 2.27 s
fib · JRuby warm: 1.84 s
fib · Spinel: 1.1 ms
Spinel 1.1 ms · JRuby warm 1.84 s · JRuby cold 2.27 s · cold = 2144× Spinel
ao_render (ray tracer): Spinel 65.4 ms; JRuby warm 2.13 s; JRuby cold 3.04 s. A cold JRuby run takes 47× the Spinel time.
ao_render (ray tracer)
ao_render (ray tracer) · JRuby cold: 3.04 s
ao_render (ray tracer) · JRuby warm: 2.13 s
ao_render (ray tracer) · Spinel: 65.4 ms
Spinel 65.4 ms · JRuby warm 2.13 s · JRuby cold 3.04 s · cold = 47× Spinel
partial_sums: Spinel 16.8 ms; JRuby warm 2.39 s; JRuby cold 2.35 s. A cold JRuby run takes 140× the Spinel time.
partial_sums
partial_sums · JRuby cold: 2.35 s
partial_sums · JRuby warm: 2.39 s
partial_sums · Spinel: 16.8 ms
Spinel 16.8 ms · JRuby warm 2.39 s · JRuby cold 2.35 s · cold = 140× Spinel
mandelbrot: Spinel 16.0 ms; JRuby warm 2.93 s; JRuby cold 1.74 s. A cold JRuby run takes 109× the Spinel time.
mandelbrot
mandelbrot · JRuby cold: 1.74 s
mandelbrot · JRuby warm: 2.93 s
mandelbrot · Spinel: 16.0 ms
Spinel 16.0 ms · JRuby warm 2.93 s · JRuby cold 1.74 s · cold = 109× Spinel
Log scale, sorted by warm JRuby time. The shaded bar spans JRuby warm (after 15 warm-up runs in one JVM) to cold (a new JVM per run); the dashed line is an empty program on a new JVM. Hover a row for its three values.
Per benchmark, log scale, sorted by warm JRuby time: Spinel, JRuby warm (same JVM, mean of the last 5 of 20 runs) and JRuby cold (a new JVM per run). The bar between warm and cold is JVM start-up plus warm-up.
Spinel: hover to isolate this series Spinel
JRuby 10.1 warm: hover to isolate this series JRuby 10.1 warm
JRuby 10.1 cold: hover to isolate this series JRuby 10.1 cold
between warm and cold: start-up + warm-up
an empty program on a new JVM: 620 ms
0.1 ms 0.1 ms
1 ms 1 ms
10 ms 10 ms
100 ms 100 ms
1 s 1 s
10 s 10 s
pidigits: Spinel 1.1 ms; JRuby warm 0.60 ms; JRuby cold 644 ms. A cold JRuby run takes 580× the Spinel time.
pidigits 1.1 ms · 0.60 ms · 644 ms
pidigits · JRuby cold: 644 ms
pidigits · JRuby warm: 0.60 ms
pidigits · Spinel: 1.1 ms
bigint_fib: Spinel 1.1 ms; JRuby warm 0.70 ms; JRuby cold 622 ms. A cold JRuby run takes 565× the Spinel time.
bigint_fib 1.1 ms · 0.70 ms · 622 ms
bigint_fib · JRuby cold: 622 ms
bigint_fib · JRuby warm: 0.70 ms
bigint_fib · Spinel: 1.1 ms
str_concat: Spinel 1.1 ms; JRuby warm 1.5 ms; JRuby cold 636 ms. A cold JRuby run takes 563× the Spinel time.
str_concat 1.1 ms · 1.5 ms · 636 ms
str_concat · JRuby cold: 636 ms
str_concat · JRuby warm: 1.5 ms
str_concat · Spinel: 1.1 ms
fannkuch: Spinel 1.1 ms; JRuby warm 13.4 ms; JRuby cold 706 ms. A cold JRuby run takes 636× the Spinel time.
fannkuch 1.1 ms · 13.4 ms · 706 ms
fannkuch · JRuby cold: 706 ms
fannkuch · JRuby warm: 13.4 ms
fannkuch · Spinel: 1.1 ms
fasta: Spinel 1.4 ms; JRuby warm 28.2 ms; JRuby cold 742 ms. A cold JRuby run takes 537× the Spinel time.
fasta 1.4 ms · 28.2 ms · 742 ms
fasta · JRuby cold: 742 ms
fasta · JRuby warm: 28.2 ms
fasta · Spinel: 1.4 ms
binary_trees: Spinel 3.2 ms; JRuby warm 50.8 ms; JRuby cold 714 ms. A cold JRuby run takes 220× the Spinel time.
binary_trees 3.2 ms · 50.8 ms · 714 ms
binary_trees · JRuby cold: 714 ms
binary_trees · JRuby warm: 50.8 ms
binary_trees · Spinel: 3.2 ms
tak: Spinel 6.5 ms; JRuby warm 62.7 ms; JRuby cold 702 ms. A cold JRuby run takes 107× the Spinel time.
tak 6.5 ms · 62.7 ms · 702 ms
tak · JRuby cold: 702 ms
tak · JRuby warm: 62.7 ms
tak · Spinel: 6.5 ms
tarai: Spinel 5.4 ms; JRuby warm 69.1 ms; JRuby cold 689 ms. A cold JRuby run takes 127× the Spinel time.
tarai 5.4 ms · 69.1 ms · 689 ms
tarai · JRuby cold: 689 ms
tarai · JRuby warm: 69.1 ms
tarai · Spinel: 5.4 ms
io_wordcount: Spinel 12.1 ms; JRuby warm 76.2 ms; JRuby cold 786 ms. A cold JRuby run takes 65× the Spinel time.
io_wordcount 12.1 ms · 76.2 ms · 786 ms
io_wordcount · JRuby cold: 786 ms
io_wordcount · JRuby warm: 76.2 ms
io_wordcount · Spinel: 12.1 ms
ackermann: Spinel 4.2 ms; JRuby warm 112 ms; JRuby cold 783 ms. A cold JRuby run takes 186× the Spinel time.
ackermann 4.2 ms · 112 ms · 783 ms
ackermann · JRuby cold: 783 ms
ackermann · JRuby warm: 112 ms
ackermann · Spinel: 4.2 ms
huffman: Spinel 4.4 ms; JRuby warm 118 ms; JRuby cold 1.08 s. A cold JRuby run takes 245× the Spinel time.
huffman 4.4 ms · 118 ms · 1.08 s
huffman · JRuby cold: 1.08 s
huffman · JRuby warm: 118 ms
huffman · Spinel: 4.4 ms
json_parse: Spinel 22.8 ms; JRuby warm 203 ms; JRuby cold 888 ms. A cold JRuby run takes 39× the Spinel time.
json_parse 22.8 ms · 203 ms · 888 ms
json_parse · JRuby cold: 888 ms
json_parse · JRuby warm: 203 ms
json_parse · Spinel: 22.8 ms
so_lists: Spinel 15.0 ms; JRuby warm 248 ms; JRuby cold 858 ms. A cold JRuby run takes 57× the Spinel time.
so_lists 15.0 ms · 248 ms · 858 ms
so_lists · JRuby cold: 858 ms
so_lists · JRuby warm: 248 ms
so_lists · Spinel: 15.0 ms
splay tree: Spinel 9.2 ms; JRuby warm 315 ms; JRuby cold 1.12 s. A cold JRuby run takes 123× the Spinel time.
splay tree 9.2 ms · 315 ms · 1.12 s
splay tree · JRuby cold: 1.12 s
splay tree · JRuby warm: 315 ms
splay tree · Spinel: 9.2 ms
linked_list: Spinel 31.3 ms; JRuby warm 331 ms; JRuby cold 992 ms. A cold JRuby run takes 32× the Spinel time.
linked_list 31.3 ms · 331 ms · 992 ms
linked_list · JRuby cold: 992 ms
linked_list · JRuby warm: 331 ms
linked_list · Spinel: 31.3 ms
template engine: Spinel 60.3 ms; JRuby warm 336 ms; JRuby cold 1.07 s. A cold JRuby run takes 18× the Spinel time.
template engine 60.3 ms · 336 ms · 1.07 s
template engine · JRuby cold: 1.07 s
template engine · JRuby warm: 336 ms
template engine · Spinel: 60.3 ms
rbtree: Spinel 18.0 ms; JRuby warm 413 ms; JRuby cold 1.03 s. A cold JRuby run takes 57× the Spinel time.
rbtree 18.0 ms · 413 ms · 1.03 s
rbtree · JRuby cold: 1.03 s
rbtree · JRuby warm: 413 ms
rbtree · Spinel: 18.0 ms
life (Game of Life): Spinel 20.8 ms; JRuby warm 430 ms; JRuby cold 1.27 s. A cold JRuby run takes 61× the Spinel time.
life (Game of Life) 20.8 ms · 430 ms · 1.27 s
life (Game of Life) · JRuby cold: 1.27 s
life (Game of Life) · JRuby warm: 430 ms
life (Game of Life) · Spinel: 20.8 ms
gcbench: Spinel 182 ms; JRuby warm 462 ms; JRuby cold 1.44 s. A cold JRuby run takes 7.9× the Spinel time.
gcbench 182 ms · 462 ms · 1.44 s
gcbench · JRuby cold: 1.44 s
gcbench · JRuby warm: 462 ms
gcbench · Spinel: 182 ms
nqueens: Spinel 4.9 ms; JRuby warm 464 ms; JRuby cold 1.41 s. A cold JRuby run takes 288× the Spinel time.
nqueens 4.9 ms · 464 ms · 1.41 s
nqueens · JRuby cold: 1.41 s
nqueens · JRuby warm: 464 ms
nqueens · Spinel: 4.9 ms
sudoku: Spinel 2.3 ms; JRuby warm 577 ms; JRuby cold 1.06 s. A cold JRuby run takes 458× the Spinel time.
sudoku 2.3 ms · 577 ms · 1.06 s
sudoku · JRuby cold: 1.06 s
sudoku · JRuby warm: 577 ms
sudoku · Spinel: 2.3 ms
matmul: Spinel 3.2 ms; JRuby warm 899 ms; JRuby cold 1.24 s. A cold JRuby run takes 389× the Spinel time.
matmul 3.2 ms · 899 ms · 1.24 s
matmul · JRuby cold: 1.24 s
matmul · JRuby warm: 899 ms
matmul · Spinel: 3.2 ms
csv_process: Spinel 98.3 ms; JRuby warm 1.19 s; JRuby cold 2.20 s. A cold JRuby run takes 22× the Spinel time.
csv_process 98.3 ms · 1.19 s · 2.20 s
csv_process · JRuby cold: 2.20 s
csv_process · JRuby warm: 1.19 s
csv_process · Spinel: 98.3 ms
sieve: Spinel 7.2 ms; JRuby warm 1.40 s; JRuby cold 1.27 s. A cold JRuby run takes 176× the Spinel time.
sieve 7.2 ms · 1.40 s · 1.27 s
sieve · JRuby cold: 1.27 s
sieve · JRuby warm: 1.40 s
sieve · Spinel: 7.2 ms
fib: Spinel 1.1 ms; JRuby warm 1.84 s; JRuby cold 2.27 s. A cold JRuby run takes 2144× the Spinel time.
fib 1.1 ms · 1.84 s · 2.27 s
fib · JRuby cold: 2.27 s
fib · JRuby warm: 1.84 s
fib · Spinel: 1.1 ms
ao_render (ray tracer): Spinel 65.4 ms; JRuby warm 2.13 s; JRuby cold 3.04 s. A cold JRuby run takes 47× the Spinel time.
ao_render (ray tracer) 65.4 ms · 2.13 s · 3.04 s
ao_render (ray tracer) · JRuby cold: 3.04 s
ao_render (ray tracer) · JRuby warm: 2.13 s
ao_render (ray tracer) · Spinel: 65.4 ms
partial_sums: Spinel 16.8 ms; JRuby warm 2.39 s; JRuby cold 2.35 s. A cold JRuby run takes 140× the Spinel time.
partial_sums 16.8 ms · 2.39 s · 2.35 s
partial_sums · JRuby cold: 2.35 s
partial_sums · JRuby warm: 2.39 s
partial_sums · Spinel: 16.8 ms
mandelbrot: Spinel 16.0 ms; JRuby warm 2.93 s; JRuby cold 1.74 s. A cold JRuby run takes 109× the Spinel time.
mandelbrot 16.0 ms · 2.93 s · 1.74 s
mandelbrot · JRuby cold: 1.74 s
mandelbrot · JRuby warm: 2.93 s
mandelbrot · Spinel: 16.0 ms
Log scale, sorted by warm JRuby time. The shaded bar spans JRuby warm (after 15 warm-up runs in one JVM) to cold (a new JVM per run); the dashed line is an empty program on a new JVM.
The JVM pays up front: it starts, loads classes, interprets, then compiles the hot methods. A warm run reuses all of that; a cold run pays it again.
After 15 warm-up runs, JRuby beats YJIT on 13 of 28 programs and Spinel on 2 (bigint_fib and pidigits, at most 0.70 ms warm): a warmed JIT re-running a tiny bignum workload, which is the best case for a long-lived process and says nothing about a cold job. On mandelbrot, sieve and partial_sums the warm mean is even slower than the cold one, so read the warm column as “JRuby after a warm-up”, not as JRuby’s peak. For JRuby on longer-running code, see the eight-fragment comparison .
Spinel has no such split: the binary starts like any C program and runs at full speed from the first instruction. For a short job (a CLI, a batch step, a helper spawned per request) start-up decides the result.
All 28 results
Benchmark Spinel CRuby 4.0 + YJIT JRuby cold JRuby warm Spinel vs YJIT Computation fib 1.1 ms 345 ms 354 ms 2.27 s 1.84 s 334× matmul 3.2 ms 239 ms 239 ms 1.24 s 899 ms 75× tak 6.5 ms 304 ms 446 ms 702 ms 62.7 ms 68× sieve 7.2 ms 380 ms 379 ms 1.27 s 1.40 s 52× mandelbrot 16.0 ms 774 ms 774 ms 1.74 s 2.93 s 48× sudoku 2.3 ms 71.1 ms 109 ms 1.06 s 577 ms 47× tarai 5.4 ms 237 ms 245 ms 689 ms 69.1 ms 45× ackermann 4.2 ms 305 ms 176 ms 783 ms 112 ms 42× partial_sums 16.8 ms 693 ms 695 ms 2.35 s 2.39 s 41× life (Game of Life) 20.8 ms 480 ms 830 ms 1.27 s 430 ms 40× nqueens 4.9 ms 145 ms 146 ms 1.41 s 464 ms 30× fannkuch 1.1 ms 8.6 ms 8.9 ms 706 ms 13.4 ms 8.0× fasta 1.4 ms 8.5 ms 9.3 ms 742 ms 28.2 ms 6.7× Data structures & GC so_lists 15.0 ms 228 ms 389 ms 858 ms 248 ms 26× rbtree 18.0 ms 267 ms 446 ms 1.03 s 413 ms 25× splay tree 9.2 ms 94.4 ms 170 ms 1.12 s 315 ms 19× gcbench 182 ms 1.78 s 2.61 s 1.44 s 462 ms 14× linked_list 31.3 ms 177 ms 336 ms 992 ms 331 ms 11× binary_trees 3.2 ms 21.2 ms 32.2 ms 714 ms 50.8 ms 9.9× huffman 4.4 ms 38.1 ms 40.0 ms 1.08 s 118 ms 9.1× Real-world programs ao_render (ray tracer) 65.4 ms 1.40 s 2.03 s 3.04 s 2.13 s 31× json_parse 22.8 ms 169 ms 275 ms 888 ms 203 ms 12× template engine 60.3 ms 433 ms 630 ms 1.07 s 336 ms 10× bigint_fib 1.1 ms 5.2 ms 5.9 ms 622 ms 0.70 ms 5.4× pidigits 1.1 ms 5.8 ms 5.9 ms 644 ms 0.60 ms 5.3× csv_process 98.3 ms 404 ms 511 ms 2.20 s 1.19 s 5.2× str_concat 1.1 ms 5.7 ms 5.6 ms 636 ms 1.5 ms 5.0× io_wordcount 12.1 ms 41.8 ms 53.4 ms 786 ms 76.2 ms 4.4×
Why Spinel can’t run a Rails app, and JRuby can
Spinel has to know every class, method and type before it emits C. Rails creates many of them after the process starts, from the database schema, the file tree and each request. Everything right of the line below is invisible to Spinel.
Build time on the left, run time on the right. JRuby carries a parser and JIT into run time; a Spinel binary has neither, so what Rails decides at boot or per request cannot reach Spinel's analysis.
BUILD TIME
binary built ▸
RUN TIME · boot → requests
JRuby: a full Ruby runtime
app + gems as Ruby source
nothing compiled ahead;
C gems swapped for Java ports
parser → interpreter → JIT
types checked while it runs
code built at boot: parsed and run ✓
Spinel: a whole-program compiler
app + libs
all source
read at once
analysis
fixes every class,
method and type
→ C → cc
native binary
classes, methods and types frozen
no parser, no interpreter inside
string eval: refused at build ✗
What Rails decides at run time
← each is an answer
the analysis needs
at build time
Spinel compiles only
what is left of the line.
ActiveRecord builds attribute methods from the schema at boot. Spinel refuses a define_method with a computed name and a module_eval of a computed string at compile time.
ActiveRecord attributes
schema → methods per column, at boot
Zeitwerk registers an autoload per file from the directory tree at boot. Spinel compiles an autoload only when its path is readable at compile time, and loads the file at program start.
Zeitwerk autoload
file tree → constants, loaded on first use
ActionView compiles each template to Ruby source at run time and evaluates it. Spinel refuses a string eval, and its ERB renders no template with Ruby code in a tag.
ActionView templates
ERB → a Ruby string → module_eval
Spinel compiles every call site to a direct C call: method_missing is never dispatched, and send with a computed name reaches only methods that exist at compile time.
method_missing · computed names
the method is picked from run-time data
Spinel cannot load a CRuby extension binary. An extension has to be rebuilt from source against Spinel's restricted C API, or bound through its FFI.
C extensions
pg · nio4r (puma) · nokogiri · bcrypt
Spinel compiles what it can see at build time. Rails computes classes, methods and code paths after boot, from the schema, the file tree and each request. Hover a row for what Spinel does with it.
Build time above the line, run time below. JRuby carries a parser and JIT into run time; a Spinel binary has neither, so what Rails decides at boot or per request cannot reach Spinel's analysis.
BUILD TIME
JRuby
Spinel
app + gems as source
nothing compiled ahead;
C gems → Java ports
app + libraries
all source read at once
analysis
fixes every class,
method and type
→ C → cc → binary
binary built ▾
RUN TIME · boot → requests
parser → interpreter → JIT
types checked while running
code built at boot:
parsed and run ✓
native binary
no parser, no interpreter
string eval:
refused at build ✗
What Rails decides at run time
ActiveRecord attributes
schema → methods per column, at boot
Zeitwerk autoload
file tree → constants, loaded on first use
ActionView templates
ERB / HAML → a Ruby string → module_eval
method_missing · computed names
the method is picked from run-time data
C extensions
pg · nio4r (puma) · nokogiri · bcrypt
↑ each is an answer the analysis needs at build time
Fix: move each decision above the line — templates to
Ruby at build, attribute methods from a dumped schema,
eager-loaded constants, ported C gems — or ship an
interpreter in the binary, which is CRuby again.
Spinel compiles what it can see at build time. Rails computes classes, methods and code paths after boot, from the schema, the file tree and each request.
Each of those is ordinary Rails, not an edge case:
ActiveRecord connects to the database at boot, reads each table’s columns and generates a reader, a writer and dirty-tracking methods per column. The class’s methods depend on the schema of the database it happened to connect to. Spinel refuses both ways of building methods at run time: a define_method with a computed name and a module_eval of a computed string.
Zeitwerk walks the autoload directories at boot, registers an autoload per file and loads each file when its constant is first used. Spinel compiles an autoload only when it can read the path at compile time.
ActionView compiles each ERB or HAML template into a Ruby source string and module_evals it on first render. Spinel’s own ERB cannot render a template with Ruby code in a tag, for the same reason: that needs eval.
method_missing and method names computed at run time. Spinel compiles every call site to a direct C call, so method_missing is never dispatched, and a send with a computed name reaches only methods that exist at compile time.
C extensions (the pg driver, puma’s nio4r, nokogiri, bcrypt) are CRuby binaries. Spinel cannot load one: an extension has to be rebuilt from source against its restricted C API, or bound through its FFI.
Making Rails compile would mean moving each decision to build time (templates rendered to Ruby, attribute methods generated from a schema dump, every constant eager-loaded, every C gem ported) or putting an interpreter in the binary, which is CRuby again.
JRuby has none of these problems because it is a complete Ruby runtime: the parser, an interpreter and a JIT are in the running process, so code that appears at boot is parsed and run like any other. The C gems are replaced with Java ports (activerecord-jdbc-adapter, nokogiri’s Java build). That is why JRuby runs Rails today and Spinel does not.
How to use Spinel from a Rails app
Spinel doesn’t need the whole app. Compile the pure, CPU-bound pieces into native binaries (“kernels”) and call them from Rails, which stays on CRuby. A kernel takes plain data in and returns plain data out: no database, no I/O beyond stdin and stdout, no metaprogramming. That is stdlib-only Ruby, which Spinel compiles today.
Rails on CRuby calls stdlib-only Ruby kernels compiled to native binaries through one KernelClient, passing plain data on stdin and stdout; CI builds the binaries for amd64 and arm64 and checks their output against CRuby's.
CRuby + Rails: unchanged
web
controllers · views · ActiveRecord · auth · the database, as today
The only code that knows kernels exist. It serialises plain data to JSON, spawns the binary for this CPU architecture, and raises on a missing binary or a non-zero exit. It never runs the kernel's Ruby source instead.
KernelClient: the one boundary
JSON → spawn binary → stdout
plain data only, no ActiveRecord objects
missing binary or exit ≠ 0 → raise
background jobs · rake tasks
batch work that calls the same
kernels through the same client
stdin: JSON
stdout: SVG / JSON
Kernels: stdlib-only Ruby, compiled by Spinel to native binaries
Good fit: projection maths and turning polygons into SVG path strings is pure CPU work on plain arrays.
map rendering
projection maths,
polygons → SVG
Good fit when the rows are already loaded: Rails does the query, the kernel only aggregates.
rollups
aggregation only;
reads stay in Rails
Good fit: numeric transforms such as value-to-colour scales and sun-angle maths have no dependencies.
colour scales
value → colour,
sun-angle maths
Good fit: parsers for text formats and batch workers that build strings from files.
parsers · batch
text formats,
string building
Build + CI
kernels/*.rb
plain Ruby source
spinel
Ruby → C → native
bin/kernels/
amd64 + arm64 builds
app image
binaries inside
deployed with the app
The kernel source is still Ruby, so CI runs every fixture through CRuby and through each compiled binary and compares the two outputs byte for byte. A difference fails the build.
CI gate
every fixture runs through CRuby and through each binary;
the two outputs must match byte for byte, or the build is red
Rails keeps the database, the views and the request cycle. Kernels get plain data and return plain data; CI proves each binary prints exactly what CRuby prints for the same input.
Rails on CRuby calls stdlib-only Ruby kernels compiled to native binaries through one KernelClient, passing plain data on stdin and stdout; CI builds the binaries for amd64 and arm64 and checks their output against CRuby's.
CRuby + Rails — unchanged
web
controllers · views · ActiveRecord · auth
background jobs · rake tasks
batch work through the same client
KernelClient — the one boundary
payload → JSON → spawn → stdout
no ActiveRecord objects, no DB handles
missing binary or exit ≠ 0 → raise
stdin: data
stdout: SVG/JSON
Kernels — stdlib Ruby → Spinel → native
one binary per arch: amd64 + arm64
map rendering
projection maths,
polygons → SVG paths
rollups
aggregation only;
reads stay in Rails
colour scales
value → colour,
sun-angle maths
parsers · batch
text formats,
string building
Build + CI
kernels/*.rb
plain Ruby source
spinel
Ruby → C → native
bin/kernels/
<name>-amd64 · -arm64
app image
binaries inside, deployed
gate
same fixture in, and the source also run on CRuby:
CRuby output == binary output
byte for byte — CRuby runs it in tests only,
else the build is red
Rails keeps the database, the views and the request cycle. Kernels get plain data and return plain data; CI proves each binary prints exactly what CRuby prints for the same input.
The rules that make it work:
Stdlib only inside a kernel. No Rails, no gems with C extensions, no database, no eval, no method_missing. That is the subset Spinel compiles today, and it also makes the code easy to test.
Plain data across the boundary. Rails loads the records, turns them into arrays and hashes and writes JSON to the kernel’s stdin. The kernel writes SVG or JSON to stdout. ActiveRecord objects and connections never cross.
One client, and it fails loudly. A missing binary, a binary for the wrong architecture or a non-zero exit raises. Rails never quietly runs the kernel’s Ruby source instead: that would hide a broken build behind a slow path that looks like it works.
CI proves the two agree. The kernel’s source is still Ruby, so CI runs each fixture through CRuby and through the compiled binary and compares the output byte for byte. A difference fails the build.
Measure first. Pick candidates by profile, not by code shape. A kernel pays only when the work is CPU time inside Ruby, and every call costs a process start (0.77 ms for an empty binary on this host) plus the JSON round trip, so batch the work per call.
The client is small:
require "open3"
class KernelClient
Error = Class . new ( StandardError )
ARCH = { "x86_64" => "amd64" , "aarch64" => "arm64" }. fetch ( RbConfig :: CONFIG [ "host_cpu" ])
DIR = Rails . root . join ( "bin/kernels" )
def self.call (name, payload)
bin = DIR . join ( " #{ name } - #{ ARCH } " )
raise Error , "kernel #{ name } is not built: #{ bin } " unless bin. executable?
out, err, status = Open3 . capture3 (bin. to_s , stdin_data: JSON . generate (payload))
raise Error , "kernel #{ name } exited #{ status. exitstatus } : #{ err } " unless status. success?
out
end
end
svg = KernelClient . call ( "map_svg" , regions: regions. map { |r| [r. id , r. polygon ] }, width: 1200 )
ARCH uses fetch, so a host the build has no binary for stops at boot instead of picking the wrong one.
Good fits: map projections and SVG generation, rollups over data already loaded, colour scales and other numeric transforms, parsers for text formats, batch workers that chew through files. Bad fits: anything bound by the database, the network or the disk (a faster loop does nothing for a request that spends its time waiting on Postgres), and anything that leans on Rails itself.
Spinel vs JRuby vs CRuby + YJIT: which to pick
Keep the app on CRuby with YJIT, use JRuby when a long-lived process needs real threads or a JVM library, and move the hot, pure pieces into Spinel kernels.
CRuby + YJIT JRuby Spinel Runs a Rails app yes, the reference yes, with JDBC adapters and Java ports of C gems no: everything must be known at build time Empty-program start 5.3 ms (ruby -e 0), 5.5 ms with YJIT 620 ms (jruby -e 0) 0.77 ms Speed vs YJIT (geomean, these 28, ARM64) 1× (interpreter: 1.2×) 0.95× after warm-up · 0.14× cold 20× Threads GVL; Ractors and fork for parallelism real threads, no GVL Thread with true parallelism and no GVL (an M:N scheduler over OS workers)String eval, method_missing, computed define_method yes yes no: refused at compile time, and method_missing is not dispatched; see its limitations C extensions yes no, Java ports instead no CRuby binaries: source rebuilt against a restricted C API, or FFI Pick it for the app itself long-lived processes that need real threads or the JVM ecosystem pure, CPU-bound kernels called from the app
FAQ
Is Spinel faster than YJIT?
On these 28 programs, yes: 20× on the geometric mean and faster on every program, from 4.4× to 334×. They are short, CPU-bound programs measured on one ARM64 host, not Rails requests.
Can Spinel run Ruby on Rails?
No. Rails builds classes and methods at run time from the schema, the file tree and its templates, and Spinel compiles only what it can see at build time. Spinel fits the pure, CPU-bound code a Rails app calls.
Is JRuby faster than CRuby?
It depends on how long the process lives. After 15 warm-up runs JRuby beats YJIT on 13 of these 28 programs (0.95× of YJIT’s speed on the geometric mean); started cold it is at 0.14×, because the JVM alone takes 620 ms to start.
Does a Spinel binary need Ruby installed?
No. The compiled executable has no interpreter or parser inside and needs only libc and libm at run time.
If you have a Rails app with a slow, CPU-bound corner, or one that needs to scale in any other way, let’s talk .
ruby