Re: Processing large amounts of files

[Liliana Marie Prikler <liliana.prikler@ist.tugraz.at>]

Am Donnerstag, dem 21.03.2024 um 16:03 +0100 schrieb Ricardo Wurmus:
> 
> Liliana Marie Prikler <liliana.prikler@ist.tugraz.at> writes:
> 
> > For comparison:
> >   time cat /tmp/meow/{0..7769}
> >   […]
> >   
> >   real  0m0,144s
> >   user  0m0,049s
> >   sys   0m0,094s
> > 
> > It takes GWL 6 times longer to compute the workflow than to create
> > the inputs in Guile, and 600 times longer than to actually execute
> > the shell command.  I think there is room for improvement :)
> 
> GWL checks if all input files exist before running the command.  Part
> of the difference you see here (takes about 2 seconds on my laptop)
> is GWL running FILE-EXISTS? on 7769 files.  This happens in prepare-
> inputs; its purpose:
> 
>   "Ensure that all files in the INPUTS-MAP alist exist and are linked
>   to the expected locations.  Pick unspecified inputs from the
>   environment.  Return either the INPUTS-MAP alist with any
>   additionally used input file names added, or raise a condition
>   containing the list of missing files."
> 
> Another significant delay is introduced by the cache mechanism, which
> computes a unique prefix based on the contents of all input files. 
> It's not unexpected that this will take a little while, but it's not
> great either.
Is there a way to speed this up?  At the very least, I'd avoid hashing
the same file twice, but perhaps we could even go further and hash the
directory once w.r.t. all the top-level inputs.  At least some
workflows would probably benefit from stratified hashing, where we
first gather all top-level inputs, then add all outputs that could be
built from those with a single process, etc.  Then again, 2 seconds vs.
10 seconds would already be a great improvement.

> The rest of the time is lost in inferior package lookups and in using
> Guix to build a script that likely already exists.  The latter is
> something that we could cache (given identical output of "guix
> describe" we could skip the computation of the process scripts).
Yeah, I think we can assume "guix describe" to be constant.  Could we
do a manifest of sorts, where the scripts are only computed once; once
we know all processes?

Cheers