1.1 KiB
Benchmarks
hello - output hello world to measure kernel calls add - run integer adds by linear fibonacci of 20 call - exercise calling by recursive fibonacci of 10 noop - a baseline that does nothing
All programs (apart from noop) run 1M times to minimize startup impact.
C was linked statically as dynamic linked influences times. Output was sent to /dev/null, so as to measure the calling and not the terminal.
Results
Results were measured by a ruby script. Mean and variance was measured until variance was low, always under one percent. Noop showed that program startup is a factor, so all programs loop to 1M.
The machine was a virtual arm (qemu) run on a acer swift 5 (i5 8265 3.9GHz), performance roughly equivalent to a raspberry pi. But results (in ms) should be seen as relative, not absolute.
language | noop | hello | add | call
c | 45 | 100 | 72 | 591
go | 53 | 4060 | 64 | 624
rubyx | 0,0374 | 1,2071 | 0,2247 | 1,3625
ruby | 1830 | 2750 | 3000 | 1900_000
Comparison with ruby, not really for speed, just to see how much leeway there is for our next layer.