RFC: Page size on PPC/PPC64 builders

David Woodhouse dwmw2 at infradead.org
Mon Mar 3 20:07:24 UTC 2008


On Mon, 2008-03-03 at 13:54 -0500, Bill Nottingham wrote:
> Josh Boyer (jwboyer at gmail.com) said: 
> > > > > For 4KiB pages:
> > > > >  - Makes life 'easier' for PPC builds, to stop people whining about it
> > > > >    finding their bugs for them.
> > > > >  - Consistent with what we actually ship in Fedora.
> > > >
> > > > - wastes less memory on lower-mem ppc machines. Not that this should be a
> > > > dealbreaker, but it's something to think about.
> > > 
> > > Yeah, that's definitely an issue. Some of the ppc boxes in our internal test 
> > > setup have "only" 1G of RAM, and when the system is under load[*], you can 
> > > run out of free pages and start oom-killing stuff pretty quick w/64k pages...
> > > 
> > > [*] http://www.codemonkey.org.uk/projects/scrashme/scrashme.git/
> > 
> > I think you're both slightly confused.  This is for switching the
> > _builders_ back to a 4KiB page kernel.  Not Fedora, which is already
> > 4KiB pages.
> 
> OK, then. Why would we explicitly make people build under a combination that isn't
> in Fedora (that they can't reproduce in Fedora) that can cause their build to fail
> (for whatever reason)? Do we appreciate being cruel?

Because it finds real bugs, and because it's useful for RHEL and EPEL that way too.

There are arguments on both sides, and I tried to list them
comprehensively in the mail which started this thread. I'm not sure
which is the 'right' answer, but it strikes me as being a little wrong
to switch back to 4KiB pages just because it's helping us find generic
bugs. If we were actually tripping over _real_ issues with 4KiB vs 64KiB
pages, that would be a different issue.

Hence the RFC.

-- 
dwmw2




More information about the fedora-devel-list mailing list