All of the hosts that will be used in the job must be included in the original hostfile - even if the hosts won't be used until comm_spawn.
The name in the hostfile must match the name provided in the info object as we otherwise would have to do DNS resolves to detect the alias, which would have a significant performance impact.
On Jan 19, 2010, at 10:05 AM, Federico Golfrè Andreasi wrote:
> Hi everyone !
> I have to submit you a question,
> I write a little code that does and MPI_Comm_spawn_multiple on a specific set of node.
> The node where to run on are specified using the "host" attribute of the MPI_Info object.
> That attribute is filled with the string get from MPI_Get_processor_name.
> Is this the correct way of working ?
> I have a strange behavior when I use alias in the /etc/hosts file.
> (I run jobs in a linux cluster, with CentOS 5).
> If I use the alias in the hostfile the spawn process throws me an error;
> instead the spawning process (run with the same hostfile) runs correctly.
> It seems that is necessary, when using the spawn function with INFO object,
> to run the spawning process with a hostfile filled with the "real" and complete hostname.
> Or I have to use a different function to get the host name?
> Thank you very much,
> users mailing list