On Fri, Feb 15, 2013 at 06:44:17PM +0000, Stuart Henderson wrote:
> On 2013/02/15 18:28, Jérémie Courrèges-Anglas wrote:
> > Stuart Henderson <[email protected]> writes:
> > 
> > [...]
> > 
> > > emacs21,no_x11 fails reliably during link:
> > 
> > Cool...
> > 
> > [...]
> > 
> > > it's pretty sad that only ~320 packages built (and several of these
> > > are just "recompress a bunch of files for things which aren't useful on
> > > vax anyway"). weird too though, there are various things which should
> > > have worked which didn't get built (e.g. math/moo). or even attempted,
> > > from the looks of the dpb logs..?
> > 
> > 320, where 5.2 shipped with 2279 packages...
> 
> yeah, something's wrong there eh? I wonder how the next build will do.
> (I'm not running them).

One known thing is that vax has crappy limits. 32M of data size, top.

Fitting the info for ~7000 ports is what killed dpb previously.  There is
a lot of code in there specific to vax, so that with -f0, dpb will be
really slow (rechecks all distfiles when building) but fit within 32M (there
is a whole lot of code in dpb and in the ports tree to try to get stuff to
vanish entirely when it's tagged with oNLY_FOR_ARCHS-*/NOT_FOR_ARCHS-*, this
shaves about 16M off dpb's memory size).

So you have it: around 5.2, a full dpb build did run up to 29M of memory
(which corresponds to the info describing the 3000 ports or so that will
be attempted...)

I've already said that this can only go so far, and it is totally wasteful
to run dpb proper on the vax itself.   dpb will be perfectly fine on a faster
machine: *everything* will be on the vax(en), including scanning the ports
tree and building things, so it's NOT cross-build.  This would free the vax'n
to do ways more useful stuff, and I'm pretty sure it will go faster (well,
less slowly is a better way to put it), as the vax is likely to swap and
thrash between dpb and what it's currently building...

dpb barely shows up as a small blip on top   when it's run on a faster
machine.  You can see it on fast machines with lots of cores, because there's
so much activity going on, but say, run dpb on amd64, connect two vaxen to
it, you will see a bit of dpb activity every hour or so, and nothing the
rest of the time. Yeah, you have to pay for two ssh connections instead of
building things locally (switching to rsh if you're on a completely
local trusted network would be a piece of cake, btw, todo list for that
very slow crap we still run onto).

(the only part of dpb that currently relies on a local ports tree is
fetching distiles and reading distinfo files, and this is on my list of
things to do... This obviously doesn't affect a dpb that runs with -f0).

Reply via email to