Re: Seeing high CPU usage during git-upload-pack

From: Nguyen Thai Ngoc Duy <pclouds@gmail.com>
Date: May 9, 2012, 10:37 UTC
Message-ID: <CACsJy8DMm0ZNiQYQ1CvOfDJu5ogP42Y06792iA_xyzeHDGLJaw@mail.gmail.com>
In-Reply-To: <CAPeUw3EsFCxki6obAAeVBb0g4MKNWBteL3VueG5djbajkFiv3A@mail.gmail.com>

On Wed, May 9, 2012 at 2:01 PM, David Ebbo <david.ebbo@gmail.com> wrote:

> We’re using git-upload-pack in a simple git server. When the repo is
> on a slow share, we’re noticing that git-upload-pack is hugging the
> CPU during the whole pack-objects operation.

What OS is it run on? What git version?

Show 13 quoted lines

> More details:
>
> - git-upload-pack gets launched (with params git-upload-pack
> --stateless-rpc //path/to/repo)
> - it itself launches a git process (with params pack-objects --revs
> --all --stdout --progress --delta-base-offset). This git process is
> the one that does all the disk I/O, but it's not the one using up the
> CPU time.
> - git-upload-pack appears to be waiting for the git process, and is at
> high CPU during that time, suggesting that it's doing some kind of
> busy wait. To further test that, I stopped the git process in the
> debugger, preventing it from making progress. At that point,
> git-upload-pack pegs the CPU forever.

How about stopping upload-pack and see where it stops? There's a main loop in upload-pack.c, create_pack_file() that is only active when data comes. I looked but failed to see how it becomes a busywait loop. Maybe you can put some debugging there.

Setting env var GIT_TRACE_PACKET and GIT_DEBUG_SEND_PACK before running upload-pack might also help.

> Would someone familiar with the sources be able to comment on this
> busy wait behavior, and on whether it could conceivably be done in a
> way that doesn’t use up some much CPU?

-- 
Duy