9 ms·

Anyone has tried modifying the go regex benchmark to use https://github.com/BurntSushi/rure-go https://github.com/BurntSushi/rure-go ? It seems to use a library

by Aissen 8y ago

Anyone has tried modifying the go regex benchmark to use https://github.com/BurntSushi/rure-go https://github.com/BurntSushi/rure-go ? It seems to use a library that isn't even on github anymore.

ysleepy 8y ago
"Go bindings to Rust's regex engine." - I't would not be in the spirit of the comparison. - Even though most scripting languages also have native regex implementations.
- burntsushi 8y ago
  The fastest Go benchmark for regex-redux is using PCRE, which is not a native Go regex engine. It's written in C.
  - Aissen 8y ago
    Precisely my point. Most scripting languages also use an engine written in C, too, as grandparent said.
burntsushi 8y ago
I had fun with this: [andrew@Cheetah benchgame] time bench-native < /tmp/input5000000.txt ... real 0m20.709s user 1m19.839s sys 0m0.183s [andrew@Cheetah benchgame] time bench-pcre < /tmp/input5000000.txt ... real 0m10.510s user 0m35.362s sys 0m0.211s [andrew@Cheetah benchgame] time bench-rust < /tmp/input5000000.txt ... real 0m2.235s user 0m3.410s sys 0m0.157s For comparison, this is the timing for the Rust program, unchanged from the benchmark site: $ time ./target/release/bench-rust-regex < /tmp/input5000000.txt ... real 0m1.452s user 0m2.643s sys 0m0.158s See https://gist.github.com/BurntSushi/9d35258444fda83de208d31fdd5e9e69 https://gist.github.com/BurntSushi/9d35258444fda83de208d31fd... for source code and full results. And no, I won't submit this myself, although I would find it absolutely hilarious if someone did and managed to get it in. :-) The PCRE version is definitely sub-optimal. I don't think it's using the JIT, for example. Enabling SIMD optimizations in Rust's regex crate shaves off another 13% (for Rust program) and 9% (for Go program). The SIMD optimizations will be enabled by default in Rust 1.27 for CPUs that support them.
- Aissen 8y ago
  Awesome work, thanks !
  - Aissen 8y ago
    This looked fun, so I did my own micro-benchmark on an rpi3, armv7 : $ for i in bench-native bench-pcre bench-pcrejit bench-rure; do echo -n $i; (cd $i; time ./$i < ../input500000.txt > /dev/null) ; echo; done bench-native real 0m24,883s user 0m56,580s sys 0m0,120s bench-pcre real 0m13,125s user 0m27,200s sys 0m0,160s bench-pcrejit real 0m2,688s user 0m3,140s sys 0m0,120s bench-rure real 0m2,779s user 0m3,300s sys 0m0,090s It looks like pcre with JIT enabled is always faster than the rure version, but this is very close.

by Aissen 8y ago

Anyone has tried modifying the go regex benchmark to use https://github.com/BurntSushi/rure-go https://github.com/BurntSushi/rure-go ? It seems to use a library that isn't even on github anymore.

ysleepy 8y ago

"Go bindings to Rust's regex engine." - I't would not be in the spirit of the comparison. - Even though most scripting languages also have native regex implementations.

burntsushi 8y ago
The fastest Go benchmark for regex-redux is using PCRE, which is not a native Go regex engine. It's written in C.
- Aissen 8y ago
  Precisely my point. Most scripting languages also use an engine written in C, too, as grandparent said.

burntsushi 8y ago

I had fun with this: [andrew@Cheetah benchgame] time bench-native < /tmp/input5000000.txt ... real 0m20.709s user 1m19.839s sys 0m0.183s [andrew@Cheetah benchgame] time bench-pcre < /tmp/input5000000.txt ... real 0m10.510s user 0m35.362s sys 0m0.211s [andrew@Cheetah benchgame] time bench-rust < /tmp/input5000000.txt ... real 0m2.235s user 0m3.410s sys 0m0.157s For comparison, this is the timing for the Rust program, unchanged from the benchmark site: $ time ./target/release/bench-rust-regex < /tmp/input5000000.txt ... real 0m1.452s user 0m2.643s sys 0m0.158s See https://gist.github.com/BurntSushi/9d35258444fda83de208d31fdd5e9e69 https://gist.github.com/BurntSushi/9d35258444fda83de208d31fd... for source code and full results. And no, I won't submit this myself, although I would find it absolutely hilarious if someone did and managed to get it in. :-) The PCRE version is definitely sub-optimal. I don't think it's using the JIT, for example. Enabling SIMD optimizations in Rust's regex crate shaves off another 13% (for Rust program) and 9% (for Go program). The SIMD optimizations will be enabled by default in Rust 1.27 for CPUs that support them.

Aissen 8y ago
Awesome work, thanks !
- Aissen 8y ago
  This looked fun, so I did my own micro-benchmark on an rpi3, armv7 : $ for i in bench-native bench-pcre bench-pcrejit bench-rure; do echo -n $i; (cd $i; time ./$i < ../input500000.txt > /dev/null) ; echo; done bench-native real 0m24,883s user 0m56,580s sys 0m0,120s bench-pcre real 0m13,125s user 0m27,200s sys 0m0,160s bench-pcrejit real 0m2,688s user 0m3,140s sys 0m0,120s bench-rure real 0m2,779s user 0m3,300s sys 0m0,090s It looks like pcre with JIT enabled is always faster than the rure version, but this is very close.