Re: The Case for Rust (in any system)

From: Alan Somers <asomers_at_freebsd.org>
Date: Fri, 13 Sep 2024 23:15:25 UTC
On Fri, Sep 13, 2024 at 1:43 PM Jason Bacon <bacon4000@gmail.com> wrote:
>
> On 9/13/24 08:21, Alan Somers wrote:
> > On Fri, Sep 13, 2024 at 7:15 AM Jason Bacon <bacon4000@gmail.com> wrote:
> >>
> >> On 9/13/24 05:33, Paul Floyd wrote:
> >>>
> >>>
> >>> On 13-09-24 06:17, David Chisnall wrote:
> >>>> On 13 Sep 2024, at 02:34, Joe Schaefer <joesuf4@gmail.com> wrote:
> >>>>>
> >>>>> I just completed a month long project to port a C++ codebase that
> >>>>> used vectors for array allocations back to using C‘s calloc. For a
> >>>>> 15% increase in memory footprint, batch jobs that took three days to
> >>>>> complete now finish in 10-12 hours.
> >>>>
> >>>> This sounds highly dubious given that std::vector is a very thin
> >>>> wrapper around malloc. From your description, I would expect the same
> >>>> speedup with some judicial use of .reserve().
> >>>
> >>> I was going to say exactly the same thing.
> >>>
> >>> Considering the reply to this, another one to be plonked so that I waste
> >>> less time.
> >>>
> >>> A+
> >>> Paul
> >>>
> >>>
> >>>
> >>
> >> Some years ago, I wrote a script to time a simple selection sort coded
> >> in various languages.  Here's an example of the results:
> >>
> >> https://github.com/outpaddling/Lang-speed/blob/master/Results/coral-amd64-100000
> >>
> >> Note: The clang array/pointer performance is currently regressed due to
> >> changes in the optimization parameters since clang 8.  That's why it's
> >> noticeably slower than GCC in these results.
> >>
> >> https://github.com/llvm/llvm-project/issues/53205#issuecomment-2318697322
> >>
> >> In general, I have not seen a significant difference between arrays and
> >> vectors in all my years running this benchmark.
> >>
> >> This benchmark is anecdotal, as it only measures performance for one
> >> algorithm.  But in my experience, C++ shows marginally slower
> >> performance and noticeably more memory use than C.
> >>
> >> Coming back to Rust: The results above, showing about double the runtime
> >> of C and C++, is the best I've seen from it.  It was taking 4x as long
> >> as C/C++ a few years ago.  That's one reason I don't use it.  I do
> >> mostly scientific computing, where runtime can be costly.  This is not
> >> *always* an issue in systems code, but it should be examined before
> >> choosing a language for a particular implementation.  The other reason
> >> is the impact of a Rust dependency on FreeBSD ports and pkgsrc packages:
> >> Frequent changes to the Rust port/package lead to long build times and
> >> frequent breakage of dependents.
> >
> > Right away, I can see that while your C program mallocs the array to
> > the list's full size, your Rust program doesn't.  It grows the list
> > one element at a time, with Vec::push .  I bet that if you change the
> > Rust program to use Vec::reserve or Vec::with_capacity, so that it
> > only has to allocate once, you'll find the results are different.
> > -Alan
>
> Thanks for the tip.  I assume this is what you meant:
>
> --- a/selsort.rs
> +++ b/selsort.rs
> @@ -19,7 +19,8 @@ fn read_list(list: &mut Vec<i32>)
>       io::stdin().read_line(&mut str).expect("failed to read input.");
>       let list_size: usize = str.trim().parse().expect("invalid input");
>       println!("list_size = {:?}", list_size);
> +    list.reserve(list_size);
>       for _c in 0..list_size
>       {
>          let mut str = String::new();
>
> It had no noticeable effect on runtime.  Not too surprising, as the vast
> majority of the runtime is in the O(N^2) selection_sort() function
> rather than the O(N) read_list().  I'd also be shocked if Vec::push()
> actually allocated 1 integer at a time.

Weird.  I checked out your repo, and for 100000 elements I get almost
the same results for Rust as for all of the C/C++ trials (but clang++
and vectors is a little bit slower).