I believe you could resolve this by specifying the interfaces to use in the
order you want them checked. In other words, you might try this:
-mca btl_tcp_if_include eth1,eth0
where eth1 is the NIC connecting the internal subnet in the cloud, and eth0 is
the NIC connecting them to the Internet. I believe OMPI will check comm in that
order, meaning that eth1 will get picked first.
Of course, that presumes something about the interfaces on your parent machine.
It doesn't matter if eth1 doesn't exist - what matters is that one of those
names is the right one to reach your cloud. If so, then this should help
resolve your problem.
On Jun 17, 2012, at 10:01 PM, Jaison Paul Mulerikkal wrote:
> HI,
>
> I'm running openmpi on Rackspace cloud over Internet using MPI_Spawn. IT
> means,
> I run the parent on my PC and the children on Rackspace cloud machines.
> Rackspace provides direct IP addresses of the machines (no NAT), that is why
> it
> is possible.
>
> Now, there is a communicator involving only the children and some
> communications
> involve only communication between children (on Rackspace cloud, in this
> scenario). When we conducted experiments, we experienced more than expected
> delays in this operation - communication between children alone.
>
> My assumption is that openMPI is looking at the direct IP addresses at the
> hostfile and try to communicate between Rackspace children over Internet.
> What I
> would want/expect is the Rackspace children communicate between themselves
> internally, using the internal Rackspace hostnames. Rackspace provide internal
> IP addresses. But if I use that in the hostfile at my home PC, the parent wont
> be able to access the children (there is a communicator involving parent and
> children).
>
> Can I anyway tell openMPI to look into the internal IP addresses of Rackspace
> machines (another hostfile, may be) for the sub-group (communicator) involving
> Rackspace children? In that case we will get performance improvement, I guess.
>
> Thanks in advance for your valuable suggestions.
>
> Jaison
> Australian National University.
>
>
> ___
> users mailing list
> us...@open-mpi.org
> http://www.open-mpi.org/mailman/listinfo.cgi/users
