gnugo-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [gnugo-devel] small plays at the end of the game


From: David G Doshay
Subject: Re: [gnugo-devel] small plays at the end of the game
Date: Sat, 4 Dec 2004 11:47:37 -0800


On 4, Dec 2004, at 6:23 AM, address@hidden wrote:

We have a few hooks into
GNU Go to maintain when you folks have a new version, and
that is all I ever see. I was trying to learn more by understanding
this problem.

Do you create GNU Go processes using fork?

No, we use MPI, Message Passing Interface.

How do you communicate with them?

We wrote a thin shim of functions around standard MPI calls,
Our code just packages things a little neater for us.

Slightly modified GNU Go is started on the master node.
Our hooks in GNU Go call SlugGo functions that result in
messages being passed to other nodes, creating a pool
manager and a pool of available workers. The master then
requests a number of workers by a message to the pool
manager. The pool manager replies with a message verifying
the workers assigned or an error if that many workers are not
available. Assigned workers block while waiting for messages
from their master. Workers may also send messages to the
pool manager to request workers reporting to them. Workers
get assigned specific jobs by their master with a message
containing the data and the job type. Results are sent back
in a message the master is waiting for. The other hook in
GNU Go is at genmove, where we trap and use our
mpi_genmove instead, and that sends us back into SlugGo
code. Similarly, SlugGo code calls back into GNU Go to do
all the work under genmove.

Cheers,
David






reply via email to

[Prev in Thread] Current Thread [Next in Thread]