I take it you are running in an rsh/ssh environment (as opposed to a managed
environment like SLURM)?
I'm afraid that you have to tell us -all- of the nodes that will be utilized
in your job at the beginning (i.e., to mpirun). This requirement is planned
to be relaxed in a later version, but that won't be out for some time.
At the moment, there is no workaround.
On 8/30/07 9:51 AM, "Murat Knecht" <MKNECHT_at_[hidden]> wrote:
> I have a question regarding the --host(file) option of mpirun. Whenever I
> try to fork a process on another node using Spawn(), I get the following
> Verify that you have mapped the allocated resources properly using the
> --host specification.
> I understand this can be fixed by providing the hostnames which will be
> used either by --host or by using a hostfile containing the names and
> possibly the slots available.
> This may be an acceptable solution, if one wants to start the same process
> on several blades, but what about starting a parent process which then
> initiates different child processes on other blades?
> In this scenario mpirun initially does not need the information of which
> other blades exist, but is only supposed to start the parent process
> locally. Surely, there must be a way not to previously specify blades, but
> to load this information at runtime, especially in a changing landscape
> where nodes are added at runtime.
> Is there a way to avoid this --host option?
> I'm using the latest version of OpenMPI (1.2.3).
> Best regards,
> users mailing list