5:44 p.m.
Hi,
simple script:
<hansl>
set verbose off
function void foo (void)
printf "Value from node %d: %d\n", $mpirank, $mpirank+1
mpisend($mpirank+1, 0)
end function
mpi --send-functions
# Call function
foo()
# Collect results
if ($mpirank == 0)
scalar Sum = 0
loop for i=0..$mpisize-1 --quiet
Sum += mpirecv(i)
endloop
printf "\nSum of received values: %d\n", Sum
endif
end mpi
</hansl>
Above code runs on Linux with either OpenMPI or MPICH, but crashes on
Windows with MS-MPI. The error code is: "DEADLOCK: attempting to send a
message to the local process without a prior matching receiv".
The problem is with sending an object (here: scalar) from root proces to
itself. When we give 2 conditions:
1. we call foo() if $mpirank !=0,
2. we receive from all nodes except root,
the code works.
The question is if this is an isse related with MS-MPI or this is gretl
specific?
Regards,
Marcin
--
Marcin Błażejowski
6:07 p.m.
New subject: Small/Serious(?) issu with MPI na MS Windows - update
On 07.10.2017 19:44, Marcin Błażejowski wrote:
Above code runs on Linux with either OpenMPI or MPICH,
The code
dosn't work on Linux with MPICH: it not crashes but gives
strange results.
Marcin
--
Marcin Błażejowski
7:24 p.m.
On Sat, 7 Oct 2017, Marcin Błażejowski wrote:
Hi,
simple script:
<hansl>
set verbose off
function void foo (void)
printf "Value from node %d: %d\n", $mpirank, $mpirank+1
mpisend($mpirank+1, 0)
end function
mpi --send-functions
# Call function
foo()
# Collect results
if ($mpirank == 0)
scalar Sum = 0
loop for i=0..$mpisize-1 --quiet
Sum += mpirecv(i)
endloop
printf "\nSum of received values: %d\n", Sum
endif
end mpi
</hansl>
Above code runs on Linux with either OpenMPI or MPICH, but crashes on
Windows with MS-MPI. The error code is: "DEADLOCK: attempting to send a
message to the local process without a prior matching receiv".
The problem is with sending an object (here: scalar) from root proces to
itself. When we give 2 conditions:
1. we call foo() if $mpirank !=0,
2. we receive from all nodes except root,
the code works.
Is an MPI_send() with dest=i matched by MPI_recv() with source=i
a well-defined operation in the MPI spec?
I can look into this, but what you're doing here looks very much
like a case for mpireduce, not send/receive.
Allin
7:57 p.m.
On Sat, 7 Oct 2017, Allin Cottrell wrote:
On Sat, 7 Oct 2017, Marcin Błażejowski wrote:
> Hi,
> simple script:
>
> <hansl>
> set verbose off
>
> function void foo (void)
> printf "Value from node %d: %d\n", $mpirank, $mpirank+1
> mpisend($mpirank+1, 0)
> end function
>
> mpi --send-functions
> # Call function
> foo()
> # Collect results
> if ($mpirank == 0)
> scalar Sum = 0
> loop for i=0..$mpisize-1 --quiet
> Sum += mpirecv(i)
> endloop
> printf "\nSum of received values: %d\n", Sum
> endif
> end mpi
> </hansl>
>
> Above code runs on Linux with either OpenMPI or MPICH, but crashes on
> Windows with MS-MPI. The error code is: "DEADLOCK: attempting to send a
> message to the local process without a prior matching receiv".
After some reading on this topic it seems you just got lucky with
OpenMPI on Linux, and in general this is something you are not
supposed to do in MPI (send/receive with the same process as
destination and source). Anyway it's pointless, just cluttering the
message space.
https://stackoverflow.com/questions/20132570/mpi-can-not-send-data-to-one...
Allin
8:08 p.m.
On Sat, 7 Oct 2017, Allin Cottrell wrote:
After some reading on this topic it seems you just got lucky with
OpenMPI on
Linux, and in general this is something you are not supposed to do in MPI
(send/receive with the same process as destination and source). Anyway it's
pointless, just cluttering the message space.
Ah, here's a more authoritative source, the MPI-3.1 documentation at
http://mpi-forum.org/docs/ , section 3.2.4:
"Source = destination is allowed, that is, a process can send a
message to itself. (However, it is unsafe to do so with the
blocking send and receive operations described above, since this may
lead to deadlock. See Section 3.5.)"
Gretl's send and receive are blocking.
Allin
8:42 a.m.
On 07.10.2017 22:08, Allin Cottrell wrote:
Gretl's send and receive are blocking.
Allin,
thanks for the answer.
In fact I did think that we can reduce only objects returned by
function, i.e. not referenced by "&" operator. As the consequence I
didn't expect that we can use more then one consecutive mpireduce().
Attached script works fine on Linux + OpenMPI or MPICH and on Windows +
MS-MPI. I suppose then it works on Mac either but I can't verify it.
Small issue: in line "mpireduce(&Sum, sum)" the word "sum" is
colorized,
is that correct?
Marcin
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
--
Marcin Błażejowski
2643
days inactive
2646
days old
5 comments
2 participants
participants (2)
-
Allin Cottrell -
Marcin Błażejowski