Re: devel/git: used one thread by default

On 2025/05/08 11:57, Stefan Sperling wrote:
> On Thu, May 08, 2025 at 09:32:00AM +0000, Klemens Nanni wrote:
> > 08.05.2025 11:33, Kirill A. Korinsky пишет:
> > > rovert@ had noticed that git out of the box uses as many threads as the
> > > system has CUPs. claudio@ had pointed that it is quite bad idea and use an
> > > example that ld.lld with -Wl,--threads=1 reduces the configure time by 15%
> > > and system time by 40%.
> > >
> > > So, here the diff which replaces 0 threads to 1 by default where it isn't 1
> > > already, and disabled try to use threads in name-hashing.
> >
> > Just ship a default config, sysmerge(8) takes care of existing files, if any.
> >
> > This is the system wide config, see git-config(1) for details.
> >
> > Feedback? OK?
>
> I prefer a config override to patched executables.
> Config files are easier to figure out for users who become curious about
> observed differences in behaviour.
>
> For years now we have been overriding configuration file defaults in the
> devel/subversion port (in that case, to disable storing plaintext passwords).
> It has never been an issue.

Trouble with a config override is that, once you've set it, if users
modify that config file, you won't get any future changes to that config
merged automatically. (There is sysmerge -p but it's extremely awkward
to use when the @sample'd config has a lot of changes because it
doesn't have the @sample contents from the time when the file was
originally changed by the user, potentially many updates ago, so it
can't do a 3-way diff).

In this particular case you can't also do "MIN(ncpu, $some_number)" with
a config override either, you have to be pessimistic and use 1, or make
things much worse on single-cpu systems.

Ignoring tests, the online_cpus value is already used in quite a few
places in the code, more than are modified by those 3 config variables
(and some don't have a config variable to control them at all afaict).

git-2.49.0/builtin/fetch.c
160: fetch_config->parallel = online_cpus();
2502: max_jobs = online_cpus();

git-2.49.0/builtin/grep.c
1194: num_threads = HAVE_THREADS ? online_cpus() : 1;

git-2.49.0/builtin/index-pack.c
2050: nr_threads = online_cpus();
2054: * max at half the number of online_cpus(), presumably because

git-2.49.0/builtin/pack-objects.c
4593: delta_search_threads = online_cpus();

git-2.49.0/name-hash.c
213: nr_cpus = online_cpus();

git-2.49.0/parallel-checkout.c
54: *num_workers = online_cpus();
63: *num_workers = online_cpus();

git-2.49.0/read-cache.c
2090: * cap the parallelism to online_cpus() threads, and we want
2280: cpus = online_cpus();
2896: cpus = online_cpus();

git-2.49.0/submodule-config.c
459: fetchjobs = online_cpus();