Re: [Libreoffice] Bug 40673 - unify build parallelization and use gnu makes --load-average feature where available

On Wed, Sep 7, 2011 at 4:39 AM, Bjoern Michaelsen
<bjoern.michaelsen@gmail.com> wrote:
> On Wed, 7 Sep 2011 04:02:36 -0500
> Norbert Thiebaud <nthiebaud-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
> wrote:
>
> > really ? ok so on my Mac I build with max-job=8 num-cpu=6... what
> > value should I use to get the same result under the new scheme?
>
> Counter question: Why would you want to get that result?

because countless run with many different values show that to a the
sweet spot on that box :-)

> max-job=8
> num-cpu=6 gets you something unpredictable between 6 and 48 jobs. The
> only way why it should work is because modern unix handles
> overcommitment quite decently (RAM might be more of an issue, but not
> with your machine). I think you would be just as fast with the new
> defaults, if not faster.
>
> > (today that means I get up to 8 jobs with up to 6 task per, except for
> > tail_build where I got 8 -- note that tail_build tend to run alone)
>
> I doubt you will ever run the 8 dmake dirs unless you compile binfilter.

all I know is that if I lower that, my elapse suffer, and I can see
that the cpu is not fully loaded... (and not I/O bound either)

> > Now, yes,  if it works (I never tested it), the load average sound
> > more promising than a pure -j n... but then if it works, then why even
> > bother with -j n at all...
> It works, but:
> a) on distributed builders (icecream, distcc) the load of the host is
>   not a good measure
> b) having no limit on -j is risky on start as make will spawn jobs like
>   mad, because most jobs are io-bound in the beginning

ok.

> c) Windows has no sensible load measure
>
> > if --with-sens-load is present then use the value specified or default
> > to nb_cpu + 1 like you said, but then use -j without value...
> see b)
>
> Also -l will lessen the pain of overcommitment: have some old dmake dirs
> running and gnu make will 'fill up' the available remaining free load
> to optimal cpu-usage.

don't get me wrong I'm all for -l :-)
my beef is with the temporary complication of the max-jobs/nb-cpu rules...
add -l support, but let's leave alone nb-cpu/max-jobs for now... a few
months and we should have almost everything it not everything under
gbuild... then things will get much simpler and saner...

> Best,
>
> Bjoern
>
> PS.:
> Well, I took a look at the count of objects in our build -- ~60% are in
> gbuild already, and of the remaining ones openssl, berkeleydb, icu,
> autodoc (which has to die), connectivity and sal are big ones. openssl
> and berkeleydb and icu are external ones which we should build with gnu
> make paralellization too.

ICU might be a problem... their build system is a home-brew nightmare.
with plenty of the dreaded 'bootstrapping' involved.  (mmeek: hints
hints :-) )

> And autodoc has to die (Did I say that
> already?).

Norbert