[webkit-dev] git.webkit.org problem

Marc-Antoine Ruel maruel at chromium.org
Thu Mar 1 06:45:25 PST 2012

Here are some things that may help;

*#1 svn server overloading when all the slaves bootstrap at the same time*
The only good work around is to keep an internal read-only svn-mirror that
the slaves use instead of the real svn server.

This means that builds must be triggered on the svn-mirror, which may
incurs a delay of a few seconds, due to svnsync polling. My hypothesis is
that svn has repo-wide locks while committing, so when there are a lot of
readers, they are all starved whenever some action is happening. Doing the
svn-mirror trick helped us to scale in the range past *several hundreds of
simultaneous active connections*. As a guidance, gclient sync does 8 svn
connections per slave by default. Scale that to hundreds of slaves waking
up at the same time...

Also, serving the read-only svn server from a set of apache frontends
helps. We do that with src.chromium.org. To be clear, above I was talking
about a DMZ-only mirror. It is a different one from the apache frontends.
Implementing both has the biggest benefit.

*#2 Using Chromium's git svn clone of webkit*
We have a few read only git frontends that are relatively speedy. I just
verified and both master at HEAD matches:
git remote add chromium http://git.chromium.org/external/Webkit.git
git fetch chromium

Once you verified that the hash matches locally with:
git fetch --all
git log -1 origin/master
git log -1 chromium/master

Then you may want to;
git config remote.chromium.fetch +refs/heads/*:refs/remotes/origin/*
git branch -r -D chromium/master

I give no guarantee what-so-ever about freshness, lack of corruption,
consistent performance, availability, YMMV, etc.
Note the url has a lower K!

Note that we also have WebKit_trimmed.git (note the upper case K!) that is
much smaller due to removal of many files chromium developers don't need
but it has invalid hashes so *do not fetch this one in your current webkit
git-svn clone*.

*#3 Buildbot's (absence of) performance*
Stefan (cc'ed) is also looking at this and has been profiling our slowest
masters for a while. In our case, we are CPU-bound and it looks like we
have to tune our caching. Stefan can explain it better than me. You guys
probably want to talk together.

Hope this is useful,


Le 29 février 2012 21:50, William Siegrist <wsiegrist at apple.com> a écrit :

> We're back to the slaves overloading svn. The Apple slaves are blocked
> temporarily while everyone else catches up. Sorry for the inconvenience.
> -Bill
> _______________________________________________
> webkit-dev mailing list
> webkit-dev at lists.webkit.org
> http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-dev/attachments/20120301/e2c168eb/attachment-0001.html>

More information about the webkit-dev mailing list