[gmx-users] re:multi problem again

[Mark Abraham]

Andrei Neamtu wrote:
> Dear Mark,
> 
> Thank you for your reply!
> 
> I thought that maybe the accessibility were the problem. I don't figure 
> how to set the working directory of say node 1 to my current working 
> directory on node 0.
> I setup node 1 to be able to access node 0 through ssh with no password 
> but the problem remains. So .. I'm stuck here.
> Any idea where to start digging solving the most probable accessibility 
> problem?

Like I said, you need to be able to see the run input file on a file 
system accessible to the other nodes.

> In short there is how I made the GROMACS installation:
> 
> 1. I installed GROMACS (serial and parallel version) on every node
> 2. I setup the node 0 to be able to access all other nodes through ssh 
> without password
> 3. generated on node 0 the nodesfile file for lam booting
> 
> I must say that parallel simulations run fine on the entire cluster (I 
> mean a single system split on all nodes). The only problem appears when 
> I try using -multi to simulate several copies of the system along the 
> nodes.

This is consistent if in a non-multi parallel program, the file I/O is 
done only on the head node and the necessary information is propagated 
across MPI. I'm not sure if this is the case in GROMACS, however.

Mark