Git Garbage Collect Error.

Hi,

Hopefully this is the right place to send bug reports... The community page "http://git-scm.com/community" suggests that it is.

      Introduction

I am creating a large GIT repository fetching from a large SVN repository, as an experiment initially. I usually use GIT repositories interfacing to parts of the SVN repository.

      Defect

During garbage collection (git gc) it encountered the following error:

git gc | git gc --prune :

    Counting objects: 856758, done.
    Delta compression using up to 2 threads.
    fatal: Out of memory, malloc failed (tried to allocate 303237121 bytes)
    error: failed to run repack

git gc --aggressive:

    Counting objects: 856758, done.
    Delta compression using up to 2 threads.
    fatal: Out of memory, malloc failed (tried to allocate 291942401 bytes)
    error: failed to run repack

At the moment the bare repository is about 4Gb in size and about 2/3rds the way through fetching.

The compression gets over 90% of the way through before this error occurs, but I don't think any compression results are kept, because when you repeat it has the same amount of work to do.

Initially this happen during an automatic gc during the fetch process. This aborted the fetch.

My system is XP64 2 core with 4Gb of memory and plenty of virtual memory.

      Comments

If this a genuine limitation due to the size of an object and memory handling limitations, then perhaps the error could be caught and the successful results kept. Ie. do a partial compression. That way the process could continue.

      Background

My requirement is to have GIT repositories of a source directory with all SVN branches included so that I can more easily merge and compare branches using GIT. However for even small source directories it takes weeks to fetch from the SVN respository (including all tags and branches), whereas fetching just the trunk takes a few hours. The SVN repository has over 90000 revisions. I am aware that I can fetch a sub-set of revisions (I don't want to at the moment), but I've found no way to fetch a sub-set of branches.

My config is as follows:

[svn-remote "svn"]
         url = svn://svn
         fetch = trunk:refs/remotes/svn/trunk
         branches = branches/*:refs/remotes/svn/*
         tags = tags/*:refs/remotes/svn/tags/*

I set this up using:

git svn init --prefix=svn/ --stdlayout --no-minimize-url svn://svn

To do this for individual directories I have to do the following:

git svn init --prefix=svn/ --stdlayout --no-minimize-url svn://svn/trunk/source/<dir>

and then edit the config manually so that:

[svn-remote "svn"]
         url = svn://svn
         fetch = trunk/source/<dir>:refs/remotes/svn/trunk
         branches = branches/*/source/<dir>:refs/remotes/svn/*
         tags = tags/*/source/<dir>:refs/remotes/svn/tags/*

This works ok but I couldn't get this result by using "git svn init" directly. Maybe I've missed something.

Regards, Tom.

On Wed, Jun 13, 2012 at 11:27:04AM +0100, Thomas Lucas wrote:

It is the right place. Sorry that you did not get any response before now.

Packing can be memory hungry if you have a lot of large objects (we may hold several large objects in memory while comparing them for deltas). It is also worse with 2 threads, as they will be working simultaneously, but in the same memory space.

Right. Nothing is written during compression; we are just coming up with a list of deltas to perform during the writing phase.

Unfortunately, I believe that the msysgit build is 32-bit, which means you are probably not even getting to use all 4Gb of your address space (my impression is that without special flags, 32-bit Windows processes are limited to 2Gb of address space).

I'd first try doing the pack single-threaded by setting the pack.threads config option to 1. If that doesn't work, you might try setting pack.windowMemory to limit the delta search based on available memory (usually it is limited by number of objects). If the large blobs are ones that do not delta well anyway (e.g., compressed media files), you might also consider setting the "-delta" attribute for them to skip delta compression entirely.

-Peff

On Thursday 12 July 2012 05:32:21 Jeff King wrote:

Indeed, that's how windows partitions memory on 32-Bit Systems. See:

http://msdn.microsoft.com/en-us/library/windows/desktop/aa366912.aspx

As it's always with that strange company, they don't spend a word about how they do it in the emulated 32 bit environment. However, a short testing reveals that a 32 bit process running on 64 bit Windows 7 with 6 GiB memory is not able to malloc() more than 1 GiB at once (which is not a big suprise at all - as malloc'ed memory has to be continuous inside the address space).

So one might guess that there is no difference in memory partitioning for 32 bit processes running on 64 bit OS.

SaCu

That's a big repo. Lots of binary files in it? Does git fsck run normally? Does it report a lot of dangling blogs/commits/etc?

Philippe