[OE-core] [RFC] Performance Issue: Build time increases

[Richard Purdie]

On Fri, 2011-08-12 at 19:06 +0800, Robert Yang wrote:
> Hi folks,
> 
> The build time of core-image-sato increases about 5 ~ 10 minutes than
> the following commit:
> 
> commit 5af197b55a4b779f1ec93186f0723026949ba2b5
> Author: Liping Ke <liping.ke at intel.com>
> Date:   Fri Jun 3 08:22:40 2011 +0800
> 
>      cache: Implement multiple extra cache fields request support
> 
> On my host,
> 1) the build time of 5af197b55a4b779f1ec93186f0723026949ba2b5 is:
> 
> real    208m26.133s
> user    241m29.280s
> sys     47m0.630s
> 
> 2) and: 068839698fe192d8846c0ed4db65861448e8e524 is:
> 
> real    217m39.687s
> user    255m34.150s
> sys     48m21.510s
> 
> 
> I use the bisect build method to find out which patch causes the time
> increases, but the build time is not stable on my host(Ubuntu 11.04 64bit),
> e.g., the build time of 1) is 208m at the first build, then "git co 
> other_commit" and build it, after about 10 builds, then go back to build
> 5af197b55a4b779f1ec, the build time will increases about 8 minutes, I have
> stopped the X and cron, at. This may have relationship with the linux
> distribution and disk.
> 
> So I have to restart the build, reboot the machine by two days, and go on the
> build. It would be better if anyone has other good method.
> 
> I think that for the next release(e.g., yocto 1.2), we can find a clean and
> stable machine(for the distribution, maybe RHEL is more stable than Ubuntu)
> to check the build time weekly, so that we can notice the performance issue
> early.

I think we really need to get to the bottom of why the build times are
so variable. Do we need to start doing these on a clean distro install?
is a fresh boot good enough? is there a way we can clear out the VM
caches and get reproducible times? reformat the build partition?

If that really is just because of the nature of a chaotic system, can we
get some better representation of build time to use as a benchmark? We
could really do with finding a faster test (which would hopefully still
be representative)...

Cheers,

Richard