From: Jeff King <peff@peff.net>
Date: Fri, 02 Oct 2026 22:19:29 GMT
Subject: Re: Windows: ~1 GB RAM per git process, many concurrent (2.56.0.windows.1)
Message-ID: <20261002221929.GB833115@coredump.intra.peff.net>
In-Reply-To: <BL0PR05MB5603A8CE8FD78127FB810BE2D8892@BL0PR05MB5603.namprd05.prod.outlook.com>

On Fri, Oct 02, 2026 at 11:04:00AM +0000, Pierre Bruno wrote:

> On Windows, git status-type commands use about 1 GB of RAM per
> process, and dozens run at once when I use coding agents (OpenCode and
> Oh My Pi) in a repository. CPU is near 0%, disk I/O is steady, and the
> total is several GB. Since two unrelated tools cause it, I suspect git
> or my repo.

Is that counting shared, mmap'd memory? Git will mmap the on-disk
packfiles (or on Windows, CreateFileMapping/MapViewOfFile). So if your
repository has a 1GB packfile, that could explain it. And it can lead to
two unintuitive conclusions:

  1. If you have several git-status processes, they're sharing all of
     those pages. So it's still only 1GB of memory use.

  2. Git will make a large map and assume the OS will fault in only the
     pages that are needed. So you may see a large virtual memory size,
     but a lower resident size (these are the terms I'd expect from
     "top" on Linux; I don't know what terms you might see on Windows).

You may also see a large resident size if the OS has faulted in all of
those pages (e.g., due to other commands). Ironically this is a sign
that you _don't_ have a lot of memory pressure (otherwise, the unused
pages would have been dropped). Again, this is coming from the Linux
side of things; I don't know how aggressively (or not) Windows is about
faulting in or releasing pages from mapped files.

-Peff