[computer-go] Pay-as-you-go cluster?

[Rémi Coulom]

Darren Cook wrote:

>I was about to suggest the first few ideas that came to mind but you've
>probably already thought of them. Would you be willing to post about the
>ideas you've already tried that failed?
>
>Darren
>_______________________________________________
>computer-go mailing list
>computer-go at computer-go.org
>http://www.computer-go.org/mailman/listinfo/computer-go/
>  
>
I have not run any experiment, but this is how I would try:

Have one master node that runs UCT and keeps the tree (or hash table) in 
memory. For each simulation, as soon as it leaves the part of the tree 
that is stored in memory, send the position over the network for 
simulation by a node. For this algorithm to work, it has to be 
asynchronous. That is to say, the master node should not wait for the 
reply of the simulation node before proceeding.

So, basically, the master node feeds the simulation nodes with positions 
from which to run simulations, and collects the results later. This 
avoids the burden of having to share the hash table over the network, at 
the price of a delay between the time when a simulation is asked and 
when its result is known. When the tree is very big, the negative 
influence of the delay should become small. If the speed of network 
communication is not high enough to keep all the simulation nodes busy, 
nodes may run more than one simulation, and/or pack simulation results 
in batches.

Of course, it is a lot easier to have such an idea than to make it 
actually work.

Rémi