[Gretl-users] Re: MPI problems
Am 10.09.2019 um 22:09 schrieb Allin Cottrell:
On Tue, 10 Sep 2019, Sven Schreiber wrote:
> Am 10.09.2019 um 20:22 schrieb Allin Cottrell:
>> Anyway, I've attempted a fix in git and today's
snapshots. If the
>> problem is not fully solved I might have to give up on real-time
>> output on Windows -- waiting for gretlmpi to finish is safer.
>
> So far that seems to have solved it, thanks very much!
Glad to hear it! I'm trying to determine if a more rigorous fix is
feasible. There are two dangers: we hang looking for additional output
from gretlmpi that's never going to arrive, or a trailing chunk of
output goes uncollected! A "rigorous fix", if possible,
would preclude both of these cases.
Hm, not sure if the current situation is right. This is still with the
Sep 10th snapshot (or should I update again?). My background problem is
that a matrix isn't written out with mwrite() from within an MPI block
(only the rank-0 process doing that), but I cannot yet produce a minimal
example.
So instead I have a question about the printout first. Consider this:
<hansl>
function void checkwritempi(void)
P = 4
mpi --send-functions
set verbose on
matrix m = I(2)
mpireduce(&m, vcat)
if $mpirank == 0
err = mwrite(m, "checkI2.mat", 1)
print err
endif
end mpi --np=P --omp-threads=1
matrix m = mread("checkI2.mat", 1)
end function
checkwritempi()
</hansl>
First of all, 'set verbose on' doesn't seem to have an effect.
Secondly, I never see output from 'print err'. In fact, the only output
I'm seeing is this:
gretlmpi 2019d-git
Aktuelle Sitzung: 2019-09-13 18:34
But at least the file checkI2.mat is written correctly...
thanks,
sven