Hi!

Kai Backman wrote:
 Hi everyone,
I put up some initial build profile results on the Wiki. The top
three time sinks are: 33% C/C++ compile, 27% dmake/build.pl,
22% dependencies (dpcc). So about 50% of our time is spent
on build or build related activities. Build time on a fast 4GB
machine is 17h.
Depends what you call fast, my slow AMD 1700XP, 512MB, reasonably
fast harddisk, needs 17h to build. ( I might be off by an hour or
so, I didn't use it lately. )

More details here:
http://wiki.services.openoffice.org/wiki/BuildSpeedup <http://www.google.com/url?sa=D&q=http%3A%2F%2Fwiki.services.openoffice.org%2Fwiki%2FBuildSpeedup> Looking at the available data the build times can probably be
decreased by a factor of 4 without source code changes. With small
source code changes the improvement could be more significant.
I'm currently doing some Jam based build tests in transex3, and
I'll tackle other parts next. Jam is a build tool written by Christopher
Seiwald and released by Perforce. They use it internally to build all
Perforce platform versions and it is used by other FOSS projects like
Boost, Freetype and Haiku. There is an old but interesting article on
transitioning from make to Jam here:
http://www.perforce.com/jam/doc/scm7.html <http://www.google.com/url?sa=D&q=http%3A%2F%2Fwww.perforce.com%2Fjam%2Fdoc%2Fscm7.html> Jam works by doing a single start up pass where it reads all the build
files and stats all sources and targets. Once the dependency graph is
ready the requested number of processes is spawned in parallel.
The source code is 12k lines of C (+2k lines in a bison parser).
Another alternative would be to change the build system to use
non-recursive dmake. The effort is probably at par with rewriting the build
system from scratch in jam. The build.pl/dmake source is 39k lines of C
and Perl. My very rough estimate for the available speedup would be 1.5-3x.
I have problems with the term "recursive dmake". Beside for dependency 
generation
there are no dmake's calling dmakes. But IMHO the main issue is that
build.pl calls dmake for each "sub-module". For example for sw/ that are 76
dmake processes that are started by build.pl and therefore 76 dmakes that
have to source the ~1 MB of makefiles in solenv/inc.

That's not recursive, but still bad.

If you would create a toplevel makefile for each module to take over the
dependency handling that build.pl currently does you are more or less
in the same situation as with jam. dmake reads the makefiles once, creates
a dependency graph and then creates the targets according to the rules.

 In any case, IMHO the best strategy would be to roll out an improved
build without replacing the existing build.pl/dmake system. Developers
can then switch once the new build system performs measurably better.
First of all: I like the idea, secondly: Nearly every new milestone breaks
the build because of some changes that worked in the Hamburg environment
but break the OOo configure based setup (aka non-official builds).
(I'm exaggerating to emphasize the problem.)

Without the support and help of the Hamburg people this will become a
maintainance nightmare.

Volker

--
If you like my work consider:  http://www.scytek.de/donations.html
PGP/GPG key  (ID: 0x9F8A785D)  available  from  wwwkeys.de.pgp.net
key-fingerprint 550D F17E B082 A3E9 F913  9E53 3D35 C9BA 9F8A 785D

Attachment: signature.asc
Description: OpenPGP digital signature

Reply via email to