Re: cross gcc-4.1.2-12 packages
On Thu, Jun 21, 2007 at 10:14:10PM -0700, Brian Morris wrote:
> I tested experimentally a little last night with this. I found using
> an old mac g3 that is roughly 15times the speed of my mac quadra,
> I get consistently about 3.5 times faster compile with simple
> examples of one page routines, no special configuration, some
> textbook code samples in c++.
Ah, nice... I've two G4s that I could use...
> watching the g3 closely with top while the quadra compiles. It looks
> like compiling a routine the mac takes about 20 seconds, three seconds
> on the g3 compiling and 2 seconds of the network (lights flash on my
> hub-switch). So I figure simple arithmetic this is about the best one can
> do. that is even if compile took 0 seconds that could only be another
> 15% improvement (or with my g4 which is 3x the g3 then 10%).
Uhm?
If the m68k mac is needing 20s and the distcc setup needs 5 sec in total,
then this is 1/4 of the time the m68k needed, or other way round 400%
faster.
During my distcc tests some years ago, I found it useable to even compile
via two DSL links (20 kB/s upstream each) between two m68ks, resulting in
50% of compile times of the standard kernel build. I don't think that
network overhead is such a problem. Think of some rather huge c++ files to
compile. Most time on m68k is spent with swapping some times. The transfer
times over network/intranet of a 10 MB c++ file is neglectable, even when
your network is just doing around 150 kB/s.
> I don't know if there could be any way to reduce this overhead factor.
The network overhead will be roughly the same, unless you can use 100 mbit
cards instead of 10 mbit ones. If the package supports concurreny_levels you
can add more distcc hosts easily.
I think we had a discussion about concurreny levels 1-2 years ago...
> in the case of packages which need a lot of configuration of course the
> benefit would be less initially but the same for recompiles on most things.
> anyway,
True. For example concurrency levels... it would be nice if the standard
Debian package building system would have such a possibility by default,
especially with the background of uprising multicore CPU coverage...
--
Ciao... // Fon: 0381-2744150
Ingo \X/ SIP: 2744150@sipgate.de
gpg pubkey: http://www.juergensmann.de/ij/public_key.asc
Reply to: