Resource efficient Thread Pools with Zig

Re:Resource efficient Thread Pools with Zig by kprotty on Monday September 13, 2021 @09:44AM (#28511380)

I'm not sure how this implies it is flawed. It benchmarks thread pools so on a system which allowed more parallel concurrent tasks (i.e. 32t amd cpu) the throughput is expected to scale somewhat, and you see this in your results.

Also, `zig -Dc` links to libc (`-Dc`) but builds in debug mode by default, similar to Rust and C/C++ compilers. The readme contains instructions to compile it with optimizations (`-Drelease-fast`), the rust versions do so as well (`--release`) so you should re-evaluate your results.

See one of my other comments in this post on why I didn't use rayon::join() by default.

[ Parent | on HN ]

Re:Resource efficient Thread Pools with Zig by pizza234 on Tuesday September 14, 2021 @09:10AM (#28524208)

> Also, `zig -Dc` links to libc (`-Dc`) but builds in debug mode by default, similar to Rust and C/C++ compilers

Right! This was a misunderstanding of mine. Updated results:

8-thread Intel laptop machine:

  zip 0.9.0-dev.958:                 230
  rust 1.5.0-nightly 21-09-12/qsort: 545
  rust 1.5.0-nightly 21-09-12/rsort: 255
  go 1.16.8:                         355

32-thread AMD workstation:

  zip 0.9.0-dev.958:                 125
  rust 1.5.0-nightly 21-09-12/qsort: 780
  rust 1.5.0-nightly 21-09-12/rsort: 135
  go 1.16.8:                         135

The reason why I think that conclusions should not be drawn due to the excessive variability between systems.

On the parent poster's system, zig is faster by a much larger margin than the two systems I've tried. And there's a ~45% difference in Go's performance compared to the fastest runner for the respective system.

[ Parent | on HN ]

Re:Resource efficient Thread Pools with Zig by kprotty on Tuesday September 14, 2021 @10:13AM (#28525137)

The results will vary on different systems given how the combination of the CPU and OS handle thread synchronization + scheduling. On one of my desktops running Windows 10 with an i7-4790k, the Go qsort runs slower (~600ms) than the Rust qsort (~490ms) while on Linux systems the results are generally the opposite.

The zig thread pool appears consistently faster for this benchmark on different systems, and rayon::join appears consistently faster than the zig thread pool on different systems too. I believe you can somewhat conclude the advantages of their designs from this relative to each other rather than in an absolute sense.

[ Parent | on HN ]