Indeed sounds like a bug in 1.6.5, but we no longer maintain that series. I'm afraid an upgrade to the 1.8 series is the only solution.
> On Mar 12, 2015, at 6:11 PM, Chris Paciorek <pacio...@stat.berkeley.edu> > wrote: > > I'm having an issue with MPI_Comm_spawn not starting workers on the > nodes provided via -machinefile or -host. This is occurring on Ubuntu > 14.04/14.10 with openMPI 1.6.5. However, I do not have the problem on > Ubuntu 12.04 with openMPI 1.4.3 nor is there a problem when I install > openMPI 1.8.4 from source on Ubuntu 14.04. > > Any suggestions as to what is going on? We can install from source to > deal with this, but Ubuntu 14.04/14.10/15.04 (and I think the current > Debian) are all relying on 1.6.5 so this issue might affect many > others. > > As far as I can tell there aren't any threads on the mailing lists or > info in the NEWS file that relate to this. > > Here's a reproducible test case. In the attached zip file, parent.cpp > generates an executable that simply tries to spawn workers using > MPI_Comm_spawn and the child executable simply reports what nodes the > workers are operating on. > > I compile and run the program as: > mpicxx -o child child.cpp > mpicxx -o parent parent.cpp > mpirun -host smeagol,arwen,smeagol,arwen -np 1 parent > > And the result is as follows with all children on the original node: > Starting: I'm process 0 and we are 1 > Finishing: I'm process 0 and we are 1 > I'm child process 0 on smeagol and we are 3 > I'm child process 1 on smeagol and we are 3 > My parent communicator size is: 3 > I'm child process 2 on smeagol and we are 3 > My parent communicator size is: 3 > My parent communicator size is: 3 > > This is all on pretty standard Ubuntu 14.04, with openMPI installed > from Ubuntu (libopenmpi-dev, libopenmpi1.6, openmpi-bin) > > I've included in the zip file: > * parent.cpp and child.cpp > * ompi_info --all on the master > * ompi_info -v ompi full --parsable on all nodes > * PATH and LD_LIBRARY_PATH info > * ifconfig information > > -Chris > > ---------------------------------------------------------------------------------------------- > Chris Paciorek > > Statistical Computing Consultant > Statistical Computing Facility, Econometrics Laboratory, Berkeley > Research Computing > > Office: 495 Evans Hall Email: pacio...@stat.berkeley.edu > Mailing Address: Voice: 510-842-9056 > Department of Statistics Fax: 510-642-7892 > 367 Evans Hall Skype: cjpaciorek > University of California, Berkeley WWW: > www.stat.berkeley.edu/~paciorek > Berkeley, CA 94720 USA Permanent forward: > pacio...@alumni.cmu.edu > _______________________________________________ > users mailing list > us...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users > Link to this post: > http://www.open-mpi.org/community/lists/users/2015/03/26463.php
