Re: [Swarm-Modelling] Question about Distributed Simulations

[Marcus G. Daniels]

address@hidden wrote:
> Therefore the communication between each node's portion of the model 
> is via the database and not from node to node.
It may be worth thinking about whether you can set up nearest neighbor 
communication.  Each node on  Blue Gene / L has six nearest neighbors, 
so if you could could directly conform the simulation to that you'd get 
higher performance.  Supposing the model is spatial, each worker node 
would have to look around it to gather (via MPI calls you'd have to 
code) relevant context for a `step' computation from up/below, 
left/right, and front/back it.
> Therefore optimizing the structure of the model in the cluster depend 
> on balancing the rate of information reading/writing into the 
> database, the computational intensity at each node, and the number of 
> nodes.  Furthermore, any computational process that can read/write 
> into the HDF database can be parallelized in this way.  Is this the 
> general gist?
Right.  It's a question of how much information really must move out of 
a local computation to a global pool in order to keep the model 
correct.  If that can be limited to about 10 MB/sec, then it looks like 
you won't have bottlenecks on this kind of system.  Latency will be much 
lower in the first approach, and bandwidth much higher, but that comes 
at the cost of low-level coding complexity in the model.   I generally 
prefer to make sure I have a problem before I fix it!

Marcus