On 11/22/2010 08:18 PM, Paul Monday (Parallel Scientific) wrote:
This is a follow-up to an earlier question, I'm trying to understand how --mca btl prioritizes it's choice for connectivity.  Going back to my original network, there are actually two networks running around.  A point to point Infiniband network that looks like this (with two fabrics):

A(port 1)(opensm)------>B
A(port 2)(opensm)------>C

The original question queried whether there was a way to avoid the problem of B and C not being able to talk to each other if I were to run

mpirun  -host A,B,C --mca btl openib,self -d /mnt/shared/apps/myapp

"At least one pair of MPI processes are unable to reach each other for
MPI communications." ...

There is an additional network though, I have an ethernet management network that connects to all nodes.  If our program retrieves the ranks from the nodes using TCP and then can shift to openib, that would be interesting and, in fact, if I run 

mpirun  -host A,B,C --mca btl openib,tcp,self -d /mnt/shared/apps/myapp

The program does, in fact, run cleanly.

But, the question I have now is does MPI "choose" to use tcp when it can find all nodes and then always use tcp, or will it fall back to openib if it can?
For MPI communications (as opposed to the ORTE communications) the library will try and pick out the most performant protocol to use for communications between two nodes.  So in your case A-B and A-C should use the openib btl and B-C should use the tcp btl.
So ... more succinctly:
Given a list of btls, such as openib,tcp,self, and a program can only broadcast on tcp but individual operations can occur over openib between nodes, will mpirun use the first interconnect that works for each operation or once it finds one that the broadcast phase works on will it use that one permanently?
If by broadcast you mean MPI_Bcast, this is actually done using point to point algorithms so the communications will happen over a mixture of IB and TCP.

If you mean something else by broadcast you'll need to clarify what you mean because there really isn't a direct use of protocol broadcasts in MPI or even ORTE to my knowledge.
And, as a follow-up, can I turn off the attempt to broadcast to touch all nodes?
See above.
Paul Monday
users mailing list

Terry D. Dontje | Principal Software Engineer
Developer Tools Engineering | +1.781.442.2631
Oracle - Performance Technologies
95 Network Drive, Burlington, MA 01803
Email terry.dontje@oracle.com