unreceived messages / MPI_Finalize

[Jianwei Li]

Joachim,

Thank you for reporting this.

You are right. We should add an MPI_Finalize() in the case of "create_file" 
(option -c in the runtime arguement) for a graceful MPI exit in the tests.
 
However, with the MPI_Finalize() fixed, I didn't get any error indicating
unreceived messages with MPICH implementation. Actually I didn't observe any 
chance for a message to hang around, since we don't use any non-blocking MPI
send/recv in our parallel-netCDF implementation. Indeed, MPI communication is
only involved in limited MPI calls in our code: MPI_Init/MPI_Finalize,
MPI_Bcast, MPI_Barrier, MPI_Allreduce, and a number of MPI_IO functions 
(currently, we don't use any non-blocking I/O in our implementation).

I hope this information could help you solve the problem soon. 

Thanks again for your interest and feedbacks.

Jianwei


>
>Hi all,
>
>I'm currently doing some tests with pnetcdf 0.93 and our MPI implementation. 
>Two things that I noticed:
>- The tests nc_test.c and nc_test.F quit w/o calling MPI_Finalize() when 
>called with '-c' (to create the test file) which results in some ugly 
>messages. Fix is of course trivial. 
>- Our MPI indicates that when MPI_Finalize() is called in the two tests above 
>(run with a single process), there may be messages which have been sent but 
>no received. Has this been observed by someone else, too?
>
> best regards, Joachim
>
>P.S: is this list archived somewhere? Couldn't find anything.
>
>-- 
>Joachim Worringen - NEC C&C research lab St.Augustin
>fon +49-2241-9252.20 - fax .99 - http://www.ccrl-nece.de