Buildfarm build_test script on an embedded device.

[Matthieu Patou]

On 05/10/2011 17:47, Christopher R. Hertel wrote:
> Some notes, as we consider how we will move forward with Embedded Samba.
>
> Matthieu's changes to the build_test script reduced the run time by only two
> hours.  That, again, emphasizes the point that it's likely the code build
> itself that is taking too long.
I was a quick shot ... without the log I can't see what's taking lot of 
time, and even with my patch you still run make test but with 1 
environment, I don't know how much time it could shave on my computer.
I would say 2 hours is not so bad.

>
> That said, I am, at this point, only using the build_test script to help get
> a sense of the problems we face working in this environment.  I believe that
> Kai is right that we need to figure out how to improve all aspects of our
> performance in the embedded space--I don't want to give this up.  Andrew is
> also correct that the build farm, as it is currently designed and working,
> is not suited to including these kinds of devices.  I also understand that
> changing the workings of the build farm to accommodate embedded platforms is
> not currently a reasonable suggestion.
>
> The solution may be to set up a separate pool of embedded devices.  I'm not
> sure yet.  I still have a lot to learn about the workings of the build farm.
I'm quite against it, I really don't see the problem of having them in 
the build farm. If the problem is basically that build is running too 
long on them and so developers who receive a mail about a breakage have 
to wait too much time before being notified to see if they broke 
something or if they fixed something I would say that it's a very bad 
excuse as basically by default nobody cares about the build farm.
I've spent enough time last year fixing and bringing it back to a 
correct state.
I remember clearly phrases of people breaking a platform that I just 
fixed and arguing that they didn't had the time to see how to fix it 
because well their stuff was very important. So no really this is *not* 
a good excuse to me. Even though, I don't see what prevents us to define 
this build machine as "we don't care" or as "only xxx cares", because 
for instance I don't care about the breakage of samba4 on Irix because 
I'm pretty sure nobody will try to run a DC on Irix.
If the speed is also a criteria, then there is a couple of machines that 
needs to be reformed as make test takes  ~10 hours to run on sun8 if you 
had at least 1 hours of compilation it's not that far from your ARM box.

I think I already had the occasion to say it but I think that for the DC 
part at least we should be able to run a make testlight (or quicktest if 
this one is appropriate) on the weakest machines so that the build and 
the test are done in a reasonable amount of time.
In the same time something has to be done to reduce the time of a make 
test because it's way too long, jelmer presented a way to display the 
time taken by each test. It's a good indicator of where there is some 
stuff to do.

Matthieu.


-- 
Matthieu Patou
Samba Team
http://samba.org