On Wed, Nov 17, 2010 at 6:55 AM, Stefan Behnel <[email protected]> wrote:
> Robert Bradshaw, 17.11.2010 09:59:
>> 1. The web site
>> 5. Buildbot
>>
>> In terms of the website itself, I think we should keep hosting that
>> (for maximum flexibility, and it's easy to administer)
>
> +1
>
>
>> and there
>> isn't much of an option for 5 (though steps should be taken to reduce
>> its load).
>
> I'm a big fan of timely commit feedback.

Me too, the build bot has been great.

> Hudson is already down to 6 build
> threads, so it takes a while now to finish the test suites. Also, the less
> parallel builds we run, the longer it disturbs other programs on the same
> machine.

Yes, though it disturbs them less. I think we'll be able to bump it
past 6 threads in the long run, it's just that things have been pretty
bad lately.

> Could you describe a little where you think the current pain points are? Is
> it more about the number of builds per day or the overall load during a
> single build? I/O or CPU load? My impression is that the /scratch/hudson
> mount point doesn't have very fast I/O, which lets the builds take longer
> than necessary.

I/O is a big pain point. /scratch/hudson isn't very fast IO, but it's
faster than the nsf mount. CPU peaks pretty badly too.

> A problem that I see is that Hudson only checks for changes every 10
> minutes in order to keep the load of the (fairly slow) Cython VM
> acceptable. So if a new commit arrives shortly after triggering the builds,
> it will continue to run the tests for 10 minutes before it starts over and
> does it all again. I tried changing that with the commit hook script but
> didn't get a response.

I looked and wasn't sure how to do it, contacted Harald Schilly (the
de-facto admin) and he wasn't sure, and then we started thinking about
migration, so this got put on hold.

The current system is particularly bad as you or I happen to push just
frequently enough to keep 12+ threads going at full tilt for the
entire time we're working on Cython.

> I don't know if bitbucket supports commit hooks, BTW.
>
> One way to get relief is to reduce the currently 9 jobs that all query the
> cython-devel repo for changes. I can try to change that by adding a
> dedicated source build job that does the polling and triggers the real
> build jobs that just download the sdist with no hg interaction. That moves
> the hg change history one step further away from the build result on the
> Hudson job page, but that won't hurt anyone, I guess.

I was actually thinking the exact same thing. Sdist is actually rather
io-intensive, but I'd imagine all jobs pulling locally (disk-to-disk)
triggered a single job that pulls from the web interface would be a
big improvement. Two other things that might help out would be (1)
doing much of the building on boxen's RAM disk and (2) having a
cascading system where, for example, 2.3 and 2.6 were fired off, and
only if all their tests passed would 2.4-2.6 run. (It seems like
there's a lot of wasteful testing going on for the non-version
specific failures.)

- Robert
_______________________________________________
Cython-dev mailing list
[email protected]
http://codespeak.net/mailman/listinfo/cython-dev

Reply via email to