Re: [gnugo-devel] using MPI for gnugo.

[Gunnar Farnebäck]

Daniel Bump wrote:
>> The question is:
>> Is it planned to use distributed parallel computing for building 
world-wide
>> AI to beat strongest human player?
>
> An immediate problem is that the GNU Go codebase is
> not well adapted to parallel processing. SlugGo is
> an experiment in using GNU Go in a parallel way
> but in order to do this right I think one would
> need to deeply hack the engine.
>
> This thread is possibly relevant:
>
> http://lists.gnu.org/archive/html/gnugo-devel/2005-04/msg00041.html
> http://lists.gnu.org/archive/html/gnugo-devel/2005-04/msg00080.html
> http://lists.gnu.org/archive/html/gnugo-devel/2005-04/msg00101.html
>
> To summarize, in 2005 Paul Pogonyshev checked in changes into
> the CVS changing the architecture to use multiple boards.
>
> The idea died because the performance hit was too high.
>
> You can get this branch from the GIT repository where it is
> called multi-board. It is a fork of GNU Go 3.7.3. If
> you clone the repository at git://git.savannah.gnu.org/gnugo.git
> you can check out the remote branch origin/multi-branch and
> rename it to a local branch with the command:
>
> git checkout -b multi-board origin/multi-board

Since then it has also become clear that Monte Carlo tree search
benefits from increased computing power a lot more than any previous
algoritms in GNU Go. The Monte Carlo code has its own board
implementation in engine/montecarlo.c and was designed with parallel
computing in mind, although there is no actual support for that in GNU
Go yet.

But before going to distributed computing I'd love to see
multithreading support of some sort in GNU Go.

/Gunnar