[distcc] Speed up linking

[Ken Beal]

One idea would be to send "partial link" requests to client machines, like
sending two object files to a participating client, which then links them
"as much as possible" and returns the resulting "combined OBJ".

It would then iterate over this process, depending on how many objects there
are to link.

Once it's down to a few "combined OBJ" files on the building machine, it
then links them together.

Lots of questions in the above, which would likely lead to configuration
variables:

1. Do we even have a tool that will "partial link"?

2. How many OBJs to send out to be linked together for best performance?

3. How many "combined OBJs" are enough to link locally for best performance?

4. It would want to be "smart" about combinations.  Perhaps instead of
combining all the OBJs, it instead sends out requests to combine each OBJ
with all the LIBs, generating a "combined OBJ" which likely has most of its
external references satisfied (so the only heavy lifting the local machine
would have to do is put together any external references from one of these
returned "combined OBJs" to another).

There may be other parts I'm not considering, but that's the benefit of
"many eyes," right?  ;-)

Sincerely,
Ken Beal


On 1/9/06, Dan Kegel <dank at kegel.com> wrote:
>
> On 1/9/06, Victor Norman <vtnpgh at yahoo.com> wrote:
> > I'm very interested in investigating the idea of speeding up the linking
> > stage.  Does anyone know of anyone doing work on this process?  Is
> > distributed and/or parallel linking even possible?
>
> I'm not sure, but I do think there are bottlenecks in the
> linker that haven't really been explored yet.  Every
> now and then somebody finds and fixes one, e.g.
> http://sourceware.org/ml/binutils/2003-09/msg00112.html
> So a sufficiently motivated developer might be able
> to speed up linking for her app by improving binutils' ld
> (and then posting the patches on the binutils mailing list
> for review).
> - Dan
>
> --
> Wine for Windows ISVs: http://kegel.com/wine/isv
> __
> distcc mailing list             http://distcc.samba.org/
> To unsubscribe or change options:
> https://lists.samba.org/mailman/listinfo/distcc
>
-------------- next part --------------
HTML attachment scrubbed and removed