[distcc] distcc is slow (was Re: survey.txt)

[Martin Pool]

On  1 Jan 2003, Simon Josefsson <jas at extundo.com> wrote:
> Martin Pool <mbp at samba.org> writes:
> 
> > Thanks for your report.
> >
> > It is strange and disappointing that it's so much slower when using
> > distcc.  What DISTCC_HOSTS setting did you use?
> 
> "hostA hostB localhost" (and "hostA localhost" for the test that
> didn't run on one platform).

I think this is the problem.  distcc favours machines that occur
earlier in the list.  You've told it to prefer the slower machines!  I
think if you put localhost first it should be better.

In particular, this will make autoconf suck because it will run many
serialized compilations remotely.

It would be nice if distcc could automatically detect and correct for
this, but it doesn't do that yet.

> > I'd be really interested to hear anything else you can discover
> > about what caused performance to be so bad.
> 
> I will try again, but given the fast localhost, slow network and low
> memory and old CPUs in the other two machines I wasn't surprised by
> the results.  Don't you ever see slowdowns when adding old machines to
> the pool?

In theory it should be possible to never have adding machines make
compilation slower.  That's not quite true now, but if you put them at
the end of the list it should be reasonably OK.

> The applications are autoconf/automake-based and doesn't use any
> makefile tweaks of any kind, so I don't think there is a dependency
> problem (unless distcc has problems with automake in general).
> 
> Thanks again for your work.
-- 
Martin