[bug#51198] [PATCH] gnu: Add b2sum.

["Ludovic Courtès" <ludo@gnu.org>]

Hi!

Leo Famulari <leo@famulari.name> skribis:

> On Wed, Oct 27, 2021 at 10:44:45AM +0200, Nicolò Balzarotti wrote:
>> This blog post [fn:1] on guix-hpc address the "Pre-built binaries
>> vs. performance" dilemma.
>>
>> [fn:1] https://hpc.guix.info/blog/2018/01/pre-built-binaries-vs-performance/
>
> It's helpful, but it doesn't look like we are using that method in Guix
> anymore. It was removed in this commit:
>
> https://git.savannah.gnu.org/cgit/guix.git/commit/?id=969adb235ee34decb65255e1ea821ff0e221ed3d
>
> I guess it learned how to do runtime feature detection?

BLIS has run-time/load-time feature detection now (which is the main
approach the blog argues for).

>> I guess the easiest way is to provide a variant (b2sum-avx or something
>> like that) with avx enabled.  Else, I'd just go with the unoptimized
>> version as it happens for many other packages, but let's hear from
>> others.
>
> The easiest thing is require local building, since it's an extremely
> cheap build. Only 1.5 seconds on my laptop, total (not just the build
> phase). With blis, one had to consider the lengthy build time.
>
> Because performance is critical for a tool like this, and because it's
> cheap to build, I've attached a patch to require local building. This is
> easier for me than creating a set of package variants that will need to
> be expanded for each new generation of CPUs :)
>
>>From 8b862425310cf631c30c823eb0fa2bfd79d36823 Mon Sep 17 00:00:00 2001
> From: Leo Famulari <leo@famulari.name>
> Date: Wed, 27 Oct 2021 11:29:40 -0400
> Subject: [PATCH] gnu: b2sum: Build on the local machine.
>
> * gnu/packages/crypto.scm (b2sum)[arguments]: Set '#:substitutable? #f'.
> ---
>  gnu/packages/crypto.scm | 13 +++++++++++--
>  1 file changed, 11 insertions(+), 2 deletions(-)
>
> diff --git a/gnu/packages/crypto.scm b/gnu/packages/crypto.scm
> index 3acd147f25..de8bcf1d70 100644
> --- a/gnu/packages/crypto.scm
> +++ b/gnu/packages/crypto.scm
> @@ -858,9 +858,18 @@ (define-public b2sum
>                   (base32 "04z631v0vzl52g73v390ask5fnzi5wg83lcjkjhpmmymaz0jn152"))))
>        (build-system gnu-build-system)
>        (arguments
> -       `(#:make-flags (list (string-append "CC=" ,(cc-for-target))
> -                            (string-append "PREFIX=" (assoc-ref %outputs "out")))
> +       `(;; By default, b2sum uses the compiler to generate instructions
> +         ;; tailored to the CPU of the running machine, using "-march=native".
> +         ;; This gives a ~1.5x speedup on a Core i5-6300U with a large dataset
> +         ;; paged in, whereas compilation of b2sum takes ~1.5 seconds.
> +         ;; b2sum does not support run-time feature detection:
> +         ;; https://github.com/BLAKE2/BLAKE2/issues/1
> +         ;; For more information, see the discussion beginning here:
> +         ;; https://debbugs.gnu.org/cgi/bugreport.cgi?bug=51198#13
> +         #:substitutable? #f

It’s okay but not entirely sufficient: on a cluster setup, you typically
talk to a daemon that’s on another machine, so the CPU features it’ll
detect there may be different from those you’ll use.  Likewise, as
discussed on IRC, you’d also need #:local-build? #t.

Anyway, I’d suggest using ‘guix hash -H blake2s-256’ or similar (it uses
libgcrypt, which does the right thing).  I think the latest Coreutils
provide a generic ‘cksum’, too, that probably does the right thing.

Ludo’.