Re: using atlocal variables in autotest macros

[John Wohlbier]

On Tue, Mar 3, 2009 at 12:14 PM, Ralf Wildenhues <address@hidden>wrote:

> * John Wohlbier wrote on Tue, Mar 03, 2009 at 05:13:17PM CET:
> >
> > I got the basic functionality to work with the following.
> >
> > m4_define([AT_MPIRUN_PROGRAM],
> > [AT_BANNER([executing mpirun -n $2 $1])
> >  AT_SETUP([$1])
> >  AT_CHECK([test -n "$MPI_LIBS" || exit 77])
> >  AT_TESTED([$1])
> >  AT_CHECK([mpirun -n $2 $1],[0],[ignore],[ignore])
> >  AT_CLEANUP
> > ])
> >
> > m4_define([AT_SERIALRUN_PROGRAM],
> > [AT_BANNER([executing ./$1])
> >  AT_SETUP([$1])
> >  AT_CHECK([test -z "$MPI_LIBS" || exit 77])
> >  AT_TESTED([$1])
> >  AT_CHECK([$1],[0],[ignore],[ignore])
> >  AT_CLEANUP
> > ])
> >
> > I was surprised by the  AT_CHECK([test -z "$MPI_LIBS" || exit 77]) line
> > (note it's different than what you suggested). I had to flip the -n and
> -z
> > tests. I guess the double pipe (||) means that if the first argument is
> > true, then take it, otherwise take exit 77.
>

Thanks again for the help.


> BTW, if I were you, then I would do this differently: don't use separate
> tests for parallel and serial, but just adjust the AT_CHECK within one
> test.
>

That seems like a reasonable suggestion. What I have a hard time seeing
though is how exactly I'd get that test to execute just once for serial, and
a test writer specified number of times with different numbers of processors
for the parallel tests. What I have now for a test is something like

$ cat src/pika_fields/test/test_Cell_Field.at
CCS_AT_SERIALRUN_PROGRAM([test_Cell_Field])

CCS_AT_MPIRUN_PROGRAM([test_Cell_Field],[1])
CCS_AT_MPIRUN_PROGRAM([test_Cell_Field],[2])
CCS_AT_MPIRUN_PROGRAM([test_Cell_Field],[4])
CCS_AT_MPIRUN_PROGRAM([test_Cell_Field],[8])

where I'd like the test author to be able to specify the numbers of
processors they want the parallel test to run with just by writing a file
such as this. I could see something like
CCS_AT_RUN_PROGRAM([test_Cell_Field],[1])
could execute serial or mpi with 1 processor based on checking $MPI_LIBS.
But I can't see how to run several tests in the case of mpi. Perhaps
something like
CCS_AT_RUN_PROGRAM([test_Cell_Field],[1],[2],[4])
I'm not sure what the AT_CHECK would look like in this case to achieve the
desired result. Are  you suggesting one AT_CHECK that handles all of the
different numbers of processors, or maybe doing a loop on the arguments
within the macro and calling an AT_CHECK per run?


> BTW2, a comment regarding portability.  The newer MPI standard specifies
> mpiexec rather than mpirun; but in practice, there are still lots of
> system differences regarding how to start MPI jobs, be that some flavor
> of mpirun, mpprun, qsub, loadlever or so.  Further, on some of the big
> iron systems, MPI is enabled and linked to by default (or even part of
> "libc"), so MPI_LIBS may as well be empty for them.  They typically have
> different program startup procedures, too.
>
>
Thanks for the heads up. I've changed mpirun to mpiexec. I think I am not
able to anticipate every system we'll run on so I'll just have to take them
as they come and try to increase portability as we go.

jgw


-- 
John G. Wohlbier