any benchmarks

[acornejo]

Found this project while looking for a c goroutine implementation. It looks wonderful! Are there any benchmarks available?

A sample implementation of the Chinese whisper benchmark would be useful to compare it against go, libtask, etc.

https://cxwangyi.wordpress.com/2012/07/29/chinese-whispers-in-racket-and-go/

[sustrik]

Have a look into "perf" subdirectory.

[acornejo]

thanks, I had looked into that folder, but didn't find that particular benchmark. Its not that I believe that particular benchmark has any inherent value, I only wanted to use it because there are implementations of it already floating online for Go, Rust, Racket, Lisp, Haskell, etc.

I just implemented one. The results are both exciting and disappointing at the same time. For values of up to 10,000, libmill seems performs well, in fact in some runs it is faster than go!

Unfortunately, with bigger values it chokes really bad. For 100000 go takes 1.7 seconds on average on my laptop, and I left the libmill version running for minutes without getting an output (so there is probably some sort of deadlock going on).

Will submit a pull-request adding the whispers benchmark to libmill.

[sustrik]

I've fixed couple of bugs and now I get this on my laptop:

sustrik@sustrik-glaptop:~/libmill$ ./perf/whispers 100000
100001
took 0.469000 seconds

Please, confirm.

[acornejo]

unfortunately, it still chokes on 100,000.

I let it run until completion, it took 164 seconds.

What operating system are you testing on? I was trying to run on an osx laptop. I just tried it on linux, and there it does complete within a second, so perhaps this is some performance bottleneck caused by clang?

[sustrik]

Unfortunately, I don't own an OSX box. For diagnostic purposes, can you check with gcc on OSX?

[acornejo]

unfortunately installing gcc in osx is a PITA.

however, I did the next best thing, I installed clang on my linux box. There I can confirm that libmill works fine with clang 3.4, so this is probably not a compiler thing.

Perhaps this has something to do with the setjmp and longjmp of glibc vs whatever libc osx uses? Or even worse, this could be related to the kernel?

I haven't looked under the hood to see how libmill actually works (may do that next weekend when I get some more spare time), so I can't offer any additional insight.

[sustrik]

My guess would be malloc() rather than set/longjmp().

Easy way to find out: Measure the setup phase (go commands) in whispers separately from message passing phase (chs/chr). If setup is slow, malloc is suspect. If sending is slow, longjmp is.