Re: [Bug-gnubg] Multiprocessing queries

[Joern Thyssen]

On Wed, Jun 04, 2003 at 03:51:49PM +0100, Ian Shaw wrote
> 
> I've been following the discussion about the multi-processing, and I
> thought I'd make a couple of comments, since I expect to make
> extensive use of it when it's ready.
> 
> I noticed there was some discussion of using hostnames rather than
> octal IP addresses. I'm on a DHCP network so the latter are liable to
> change.

Eventually you'll be able to use either. 

> Will the "master" be able to tell when new "slaves" are available? 

I'll just answer the questions to my best knowledge. Olivier, please
correct me if I wrote something wrong :-)

No, in the current implementation you'll have to tell the master that a
slave is available with a command like

pu add remote 1.2.3.4:10000

to tell the master that a slave is ready at IP address 1.2.3.4 listening
on port 10000. The master will then contact the slave, exchange a hand
shake, and they're ready to rock'n'roll :-)

The slaves will answer and server any master (subject to a network
mask), so they don't know who will be their master, thus they cannot
notify any. Possible solutions are: (a) send a 0.0.0.0 broadcast to all
computers on the network. Masters should return a reply that they would
like the slave to work for them, or (b) allow the user to specify a
default master; when the slave becomes available it'll notify the
default master. If the master says "go away" it'll just start listening
on the specified port, and wait for any other master to contact them.

> I am likely to be able to start a rollout, then later set up more
> processors when my colleagues go home.  Will it cope with slaves being
> switched off mid-operation? 

I don't know. Olivier will have answer that one. The code seems to be
well designed and coded, so I'll attempt a guess and say: "yes!".

Anyway, you can't add slaves while doing a rollout. You'll have to add
them when the current rollout has finished.

Jørn