I've just got the Intel Cluster Studio XE 2013 fro Linux OS. And I installed it on the current machine (the master node). After that, I let other nodes in the cluster mount the /opt of the master node through nfs. The mpdboot could work well. But when I used mpirun to run the program, I found that all the processes were running on the master node. If I use -hosts to force the program to run on a specific node other than the master one, it got stuck and no output would be generated. I was wondering if it was because that I didn't use distribution installation to install it on all the nodes. Isn't it right to install it on a single machine and the other machines just use it through nfs? I am sure that there is nothing wrong with the program. Because we used other mpi and even the old version of the mpi, it worked well.
Any one can give me some hints? Thanks!
Not only the MPI libraries but your application must be visible on all nodes on the PATH which is set up in a new shell.
You could take care of the MPI libraries with an NFS mount and sourcing the mpivars (and probably compilervars) in each process, but you would also have to take care that each process gets into a working directory.