[FLASH-USERS] Running simulations on a cluster

[Seyit Hocuk]

Dear Flash-users Community,


I have a problem which I heard before and am experience it now myself. I 
hope it was solved and that people can help me with it.

I am resolution limited while running simulations on 1 cpu, however 
whenever I use the computer cluster that is available to me with 20 
cpu's, I notice that it's not faster at all. In fact, it's even a little 
bit slower. My machine is not that great compared to the machines of the 
cluster.

I suspect that dividing the tasks is not done very well. Perhaps it is 
some kind of optimalization problems, because I use an intel cpu and the 
cluster has amd (Opteron). Or it could be that I/O is very slow on the 
cluster, but this should not be the limiting case. I am afraid though 
that my setting up of the simulation might be wrong, because every node 
(machine) has to connect to the master cpu at every timestep (dt).

Any help is much appreciated.

Kind regards,
Seyit


Btw: I am using mpich2 and create a ring of cpu's with mpdboot 
--totalnum=.. --file=.. and start the simulation with mpiexec -n X. With 
mpdtrace I do see that multiple machines are within the ring and so 
directly assume that multiple machines are used. I don't know any other 
way to see that the task is divided.