[mvapich-discuss] NaNs from non-blocking comms

[Sayantan Sur]

Hi Dan,

Thanks for posting this example. I took a quick look at the example. I
think there is a bug in the application code. The MPI standard
requires that all non-blocking communications be (locally) completed
before calling finalize. MPI_Barrier doesn't guarantee this. Let me
know if you think I am mistaken.

http://www.mpi-forum.org/docs/mpi-2.2/mpi22-report.pdf

Page 291, line 36

"This routine cleans up all MPI state. Each process must call
MPI_FINALIZE before
it exits. Unless there has been a call to MPI_ABORT, each process must
ensure that all
pending nonblocking communications are (locally) complete before
calling MPI_FINALIZE."

Can you try inserting MPI_Wait / MPI_Waitall in your example to see if
this works?

Thanks!

On Tue, Apr 5, 2011 at 10:59 AM, Dan Kokron <daniel.kokron at nasa.gov> wrote:
> Using mvapich2-1.6 configured and built under x86_64 Linux with
>
> Intel-11.0.083 suite of compilers
>
> ./configure CC=icc CXX=icpc F77=ifort F90=ifort CFLAGS=-fpic
> CXXFLAGS=-fpic FFLAGS=-fpic F90FLAGS=-fpic
> --prefix=/home/dkokron/play/mvapich2-1.6/install/intel --enable-f77
> --enable-f90 --enable-cxx --enable-romio --with-hwloc
>
> The attached example code gives NaN's as output from the MPI_Recv if
> MV2_ON_DEMAND_THRESHOLD is set to be less than the number of processes
> used.
>
> The example also gives NaNs using IntelMPI-4.0.1.002 if
> I_MPI_USE_DYNAMIC_CONNECTIONS=enable
>
> See the 'commands' file in the tarball for more information.
> --
> Dan Kokron
> Global Modeling and Assimilation Office
> NASA Goddard Space Flight Center
> Greenbelt, MD 20771
> Daniel.S.Kokron at nasa.gov
> Phone: (301) 614-5192
> Fax:   (301) 614-5304
>
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>
>



-- 
Sayantan Sur

Research Scientist
Department of Computer Science
http://www.cse.ohio-state.edu/~surs