Thanks!† That worked.


From: [] On Behalf Of Ralph Castain
Sent: Wednesday, April 17, 2013 7:45 AM
To: Open MPI Users
Subject: [External] Re: [OMPI users] Change In Behavior Running on Multiple Nodes


Hmmm...yes, I see what you mean. Try using "-default-hostfile mpimachines" instead of -hostfiles. I think that's the problem - at least, it fixed it for me.



On Apr 17, 2013, at 7:08 AM, "Welder, Wallace T         PWR" <> wrote:

Iíve recently tried running OpenMPI v1.6.4 on multiple nodes and have noticed a change in behavior that I donít understand.  In OpenMPI version 1.4.x, 1.5.x and 1.6.1, I could run a job spanning two nodes as shown below.  The procedure results in 8 processes running on the first node and 8 on the second node.

mpirun Ėhostfile mpimachines Ėn 1 host.exe : -n 15 node.exe


where the file mpimachines looks  like:

node1 slots=8

node2 slots=8

In OpenMPI v1.6.2 and v1.6.4 (havenít tried v1.6.3) when I try to run the same way, all the processes start on node1 and none start on node2.  Iíve noticed there are now runtime flags -bynode and Ėbyslot, but  I havenít had any success with those.  Iíve also tried changing the mpimachines file to look like:

node1 slots=8 max-slots=8

node2 slots=8 max-slots=8

When I tried this, I got a runtime error saying there were not enough slots in the system to satisfy the 15 slots that were requested by the application node.exe.  I think there is a hint in here about my problem, but I havenít been able to figure out what it is yet.

Can anyone let me know how the process has changed with these newer versions of OpenMPI?






users mailing list