[fedora-arm] Who's using Kirkwood?

[Brendan Conoboy]

On 10/11/2012 03:10 AM, Gordan Bobic wrote:
> Just out of interest, which packages are you referring to? I am assuming
> it is LibreOffice + a small subset of whatever is in Fedora that isn't
> in EL; mainly because I had no RAM/swap/CPU issues building any the 2000
> or so packages that overlap. Takes about 3-4 weeks on a _single_
> SheevaPlug.

You're building 2000 packages, we're building 12000.  Libreoffice is 
definitely one one of the problem packages where an armv7hl builder is 
called for.  The koji server has a special 'heavybuilder' group which 
handles such packages.  Are you using USB storage on your sheevaplug? 
It surprises me that you can get through even 2000 in 3 weeks unless 
half of them are noarch ;-)

>> 3. Certain features such as atomic operations aren't available on armv5,
>> reducing the number of packages that can be built for ARM in total: If
>> it fails on armv5 but works on armv7, we still don't get it for armv7.
>
> In _most_ packages that require this, there are patches that address it.

According to 
http://fedoraproject.org/wiki/Architectures/ARM/Fedora17_rawhide 
openmpi, pixie, mongodb are all currently broken due to atomics.  This 
blocks condor, iwhd, perl-MongoDB, netcdf*, espresso, gdl, gdal, 
gromacs, ScientificPython, towhee, pypar, orsa, R-RScaLAPACK, nco, which 
in turn blocks even more packages.  This is not an exhaustive list. 
This also doesn't consider that some package builds are transiently 
successful and transiently fail due to thread-safe issues which aren't 
coded for in armv5tel.  With 12000 packages you never known when an 
armv5tel build is going to hit an SMP builder and expose such a bug.  It 
happens all the time, but koji-shadow just reissues these builds so they 
work on a subsequent build... sometimes.  Or they block hundreds of 
packages because of a transient failure.

>> 5. On the whole, it's not a popular Fedora ARM target. Raspberry pi,
>> OMAP, highbank, this is where most (not all) of our known users have
>> hardware and interest. There are some Kirkwood users, clearly, but there
>> are a lot more users of everything else. We should get some updated
>> download stats on this to demonstrate, but last I saw kirkwood was maybe
>> 3% of usage.
>
> Perhaps a poll might be a good way to ascertain this, rather than a
> discussion?

Feel free to organize one, but what will you do with the resulting data? 
  Regardless of the result, the bottom line is that the people who 
volunteer to do the work get to set the direction.  The plug devices are 
perfectly useful, still in production, but there just isn't enough 
manpower interested and capable of supporting them over the long term. 
The way to keep kirkwood alive isn't to justify its existence, it's to 
do the work to keep it running.

-- 
Brendan Conoboy / Red Hat, Inc. / blc at redhat.com