threads / discuss / 64085

Performance (self-compiled vs. distributed)

Subject: Performance (self-compiled vs. distributed)

## tl;dr

3 messages between Sep 3, 2025 and Sep 4, 2025.

replies: 2people: 1as markdown or json

D. Ben Knoble· Sep 3, 2025, 20:31 UTC · lore

I have an installation of Git at /usr/local/bin/git managed by Homebrew. I'm also tinkering with building my own Git (+ patches) to experiment with new features. Today I used "make install …" to get that in ~/bin/git, but I'm noticing some performance problems. My shell prompt started coming back slower, especially in a large monorepo.

For example, inside said monorepo (apologies in advance; output is on a computer I can't easily copy from):

    hyperfine -Ni -w10 {/usr/local,$HOME}/bin/git
    …
    /usr/local/bin/git ran 4.99 +/- 1.69 times faster than ~/bin/git

The difference is 22ms +/- 6.5ms (15.3ms–56.3ms) compared to 109.7ms +/- 17.8ms (91.1ms–159.9ms).

My build recipe was

    export DEVELOPER=1
    export XML_CATALOG_FILES=/usr/local/etc/xml/catalog
    make -j "$(nproc)" all doc
    make install install-doc install-html

Which differs somewhat from Homebrew's recipe (namely, Perl & Python path aren't set; nor is USE_LIBPCRE nor INSTALL_SYMLINKS… I'm ignoring the "extras" for now). I wouldn't particularly expect any of those to create an issue, though.

Any tips on debugging this would be greatly appreciated.

Some other tidbits:

    ls -lh@ {/usr/local,$HOME}/bin/git
    -rwxr-xr-x@ … 4.4M … ~/bin/git
        com.apple.provenance     11B
    -r-xr-xr-x@ … 3.5M … /usr/local/bin/git
        com.apple.provenance     11B

I wondered if quarantining was at play, and this seems to indicate it isn't. It is interesting that my binary is so much bigger!

Source I compiled: https://github.com/benknoble/git/tree/ps-jj

PS I also noticed we put all the RelNotes in $prefix/share/doc/git-doc/RelNotes/*.adoc; I might have expected those to get turned into HTML, too?

-- 
D. Ben Knoble

D. Ben Knoble· Sep 4, 2025, 14:05 UTC · re: D. Ben Knoble · lore

Re: Performance (self-compiled vs. distributed)

On Wed, Sep 3, 2025 at 4:31 PM D. Ben Knoble <ben.knoble@gmail.com> wrote:

Show 24 quoted lines

>
> I have an installation of Git at /usr/local/bin/git managed by
> Homebrew. I'm also tinkering with building my own Git (+ patches) to
> experiment with new features. Today I used "make install …" to get
> that in ~/bin/git, but I'm noticing some performance problems. My
> shell prompt started coming back slower, especially in a large
> monorepo.
>
> For example, inside said monorepo (apologies in advance; output is on
> a computer I can't easily copy from):
>
>     hyperfine -Ni -w10 {/usr/local,$HOME}/bin/git
>     …
>     /usr/local/bin/git ran 4.99 +/- 1.69 times faster than ~/bin/git
>
> The difference is 22ms +/- 6.5ms (15.3ms–56.3ms) compared to 109.7ms
> +/- 17.8ms (91.1ms–159.9ms).
>
> My build recipe was
>
>     export DEVELOPER=1
>     export XML_CATALOG_FILES=/usr/local/etc/xml/catalog
>     make -j "$(nproc)" all doc
>     make install install-doc install-html

Interesting. It was suggested off-list that I should try building with CFLAGS=-Oz, which I'm having some trouble propagating (with V=1, I see some initial files built with -Oz and then later some with -O2??).

But on my other (non-corporate-managed) machine, I see almost no timing difference between a built Git and a Homebrew-installed Git. So I wonder if the corporate binary sniffer is causing problems…

-- 
D. Ben Knoble

Ben Knoble· Sep 4, 2025, 21:59 UTC · re: D. Ben Knoble · lore

Re: Performance (self-compiled vs. distributed)

Show 31 quoted lines

> Le 4 sept. 2025 à 10:05, D. Ben Knoble <ben.knoble@gmail.com> a écrit :
> 
> On Wed, Sep 3, 2025 at 4:31 PM D. Ben Knoble <ben.knoble@gmail.com> wrote:
>> 
>> I have an installation of Git at /usr/local/bin/git managed by
>> Homebrew. I'm also tinkering with building my own Git (+ patches) to
>> experiment with new features. Today I used "make install …" to get
>> that in ~/bin/git, but I'm noticing some performance problems. My
>> shell prompt started coming back slower, especially in a large
>> monorepo.
>> 
>> For example, inside said monorepo (apologies in advance; output is on
>> a computer I can't easily copy from):
>> 
>>    hyperfine -Ni -w10 {/usr/local,$HOME}/bin/git
>>    …
>>    /usr/local/bin/git ran 4.99 +/- 1.69 times faster than ~/bin/git
>> 
>> The difference is 22ms +/- 6.5ms (15.3ms–56.3ms) compared to 109.7ms
>> +/- 17.8ms (91.1ms–159.9ms).
>> 
>> My build recipe was
>> 
>>    export DEVELOPER=1
>>    export XML_CATALOG_FILES=/usr/local/etc/xml/catalog
>>    make -j "$(nproc)" all doc
>>    make install install-doc install-html
> 
> Interesting. It was suggested off-list that I should try building with
> CFLAGS=-Oz, which I'm having some trouble propagating (with V=1, I see
> some initial files built with -Oz and then later some with -O2??)

Fixed this, which helped with the binary size…

> But on my other (non-corporate-managed) machine, I see almost no
> timing difference between a built Git and a Homebrew-installed Git. So
> I wonder if the corporate binary sniffer is causing problems…

…but not speed. Sigh. Assuming I can rule in or out the corporate parts, I’ll keep troubleshooting. Might try to install to the Homebrew paths just to see what happens. (I’m assuming the sniffer uses a fingerprint though and not just the path.)