[distcc] about distcc's scheduling

[Francois Berenger]

Fergus Henderson wrote:
> Have you tried using pump mode?  That helps a lot.
> 
> Also, if the issue is that distcc isn't running enough jobs on each 
> host, try using the "host/JOBS" format in DISTCC_HOSTS,
> e.g. DISTCC_HOSTS="foo.riken.jp/20,cpp,lzo 
> <http://foo.riken.jp/20,cpp,lzo> bar.riken.jp/40,cpp,lzo 
> <http://bar.riken.jp/40,cpp,lzo>".
> This would allow up to 20 concurrent compiles on host foo.riken.jp 
> <http://foo.riken.jp> and 40 concurrent compiles on host bar.riken.jp 
> <http://bar.riken.jp>.
> (The ",cpp,lzo" part is required for pump mode.)

Thanks a lot! I am trying this right away.
I'll post my speedup results when I'm done.

> At some point, depending on the size and structure of the software that 
> you're trying to compile,
> and the way your Makefiles are arranged, you may run into limitations in 
> the parallelizability of your build graph.
> For example, if your software has only three source files, then distcc 
> won't be able to speed up compilation by more than a factor of three at 
> very best.
> Similarly, if your software has 30 files, but they are split into 10 
> directories of three files each, and your Makefile(s) sequentialize the 
> compilation of each directory, then you still won't get more than a 
> factor of three speed up at very best.

We use scons. I am not so happy with it (like when it spends 30s before 
giving me back control of the command line when theres is _nothing_ to 
compile because everything is already done).
However, it manages the dependencies far better than any Makefile I 
could have written by hand.

Regards,
Francois.

> On Mon, Nov 23, 2009 at 8:19 PM, Francois Berenger <berenger at riken.jp 
> <mailto:berenger at riken.jp>> wrote:
> 
>     Hello,
> 
>     Is there a way to use a pull scheduler instead of a push scheduler
>     inside distcc?
> 
>     I have many compiling hosts and when they join a compiling job, they are
>     not enough loaded, at least to my taste. Meaning: they could compile
>     more things and I could wait less for them to finish.
> 
>     I think the problem is that the scheduler is pushing: the distcc
>     server sends jobs to distcc compiling hosts.
> 
>     Instead, I would like this behavior: distcc compiling hosts ask and
>     take jobs as long as they are not completely loaded with compilation
>     jobs.
>     The distcc server should distribute jobs to whom asks for.
> 
>     This should scale better.
> 
>     If there is no way now to do so but you can give me some hint on how to
>     do so (where to look, what to modify, etc.), it might be helpful.
> 
>     Best regards,
>     Francois.
> 
>     __
>     distcc mailing list            http://distcc.samba.org/
>     To unsubscribe or change options:
>     https://lists.samba.org/mailman/listinfo/distcc
> 
> 
> 
> 
> -- 
> Fergus Henderson <fergus at google.com <mailto:fergus at google.com>>