Re: serious performance issues with images, audio files, and other "non-code" data

On Fri, May 14, 2010 at 08:54:02AM -0400, John wrote:

Thanks so much. It's version 1.5.6.5. I compiled it 3 months ago. For

By git standards, that version is ancient. You may want to try with a
more recent version of git (at the very least, multithreaded delta
compression has been enabled by default since then).

I packed the bare repo, then ran `gc --aggressive`.

Note that "gc --aggressive" will repack from scratch, throwing away the
previous pack.

Then I did a `git pull`, which took 35 minutes.

That sounds like a long time. What was taking so long? Was delta
compression pegging the CPU? Was it limited during the "Writing objects"
phase, which is going to be limited by either disk I/O or network speed?

How big is your packed repo? Given the pattern you describe below, I am
beginning to wonder if it is simply the case that even though a single
checkout of your repo isn't that large, the complete history of your
project may simply be gigantic (e.g., because you are repeatedly writing
new apparently-random versions of each file, so your repository size
will grow quite quickly).

Remember that a git clone transfers the full history (and a pull will
transfer all of the intermediate history). If you have rewritten those
files many times, you may be transferring many times your working
directory size in history.

You can simulate it all by generating a batch of 1-100 MB files from
/dev/urandom (since they won't compress), commit them, then do it
again many times to simulate edits. Every few iterates, push it
somewhere.

I tried this script to make a 100M working directory with a 400M .git
directory:

-- >8 --
#!/bin/sh

rm -rf big-repo
mkdir big-repo && cd big-repo && git init

mark() {
  echo "`date` $*"
}

randomize() {
  mark randomize start
  for i in `seq 1 100`; do
    openssl rand $((1024*1024)) >$i.rand
  done
  mark randomize end
}

commit() {
  mark add start
  git add .
  mark add end
  mark commit start
  git commit -m "$1"
  mark commit end
}

randomize; commit base
randomize; commit one
randomize; commit two
randomize; commit three
-- 8< --

Here are a few timings I noted:

  - it takes about 5 seconds to generate and write the random data

  - git add runs in about 13 seconds. It pegs the CPU hashing all of the
    data.

  - the first commit is nearly instantaneous, as the summary diff takes
    no work; subsequent commits spend about 9 seconds to create the
    summary diff.  Changing commit to "commit -q" drops that to back to
    near-instantaneous.

  - with no attributes set, "time git gc --aggressive" reports:

      real    1m31.983s
      user    2m29.621s
      sys     0m3.732s

    Note the real/user discrepancy. It's a dual-core machine, and recent
    git will multi-thread the delta phase, which is what dominates the
    time. This should correspond roughly to the delta-compression phase
    of your pull time, as that was just making a pack on the fly (but
    now that we are packed, pulls will be limited only by the time to
    transfer the objects themselves).

  - Turning off delta compression for the .rand files makes repacking
    much faster:

      $ echo '*.rand -delta' >.gitattributes
      $ time git gc --aggressive
      ...
      real    0m25.354s
      user    0m22.057s
      sys     0m1.316s

    The delta compression phase is very quick, and we spend most of our
    time writing out the packfile to disk.

So I stand by my earlier statements:

  1. Use "git commit -q" to avoid wasting time on the commit diff
     summary (we should perhaps have a commit.quiet config option for
     repos like this where you would almost always want to suppress it).

  2. Make sure your upstream repo is packed so pullers do not have to
     generate a new packfile all the time.

  3. Use -delta where appropriate to avoid useless delta compression.

If things are still slow after that, you'll need to be more specific
about your exact workload and exactly what is slow (I am still not sure
if delta compression or network bandwidth is the limiting factor for
your slow pulls).

-Peff