[OMPI users] MPI_SCAN

2012-05-15 Thread Rohan Deshpande
I am performing Prefix scan operation on cluster I have 3 MPI tasks and master task is responsible for distributing the data Now, each task calculates sum of its own part of array using GPUs and returns the results to master task. Master task also calculates its own part of array using GPU. Whe

Re: [OMPI users] problems with 1.6

2012-05-15 Thread David Turner
Hi Ralph, Sorry for the false alarm, and thanks for the tip: ... version confusion where the mpirun being used doesn't match the backend daemons. Yes, my test environment was wonky. All is well now. On May 14, 2012, at 3:41 PM, David Turner wrote: ... [c0667:24962] [[39579,1],11] ORTE_E

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Jeff Squyres
On May 15, 2012, at 4:13 PM, Ricardo Reis wrote: > printing the result of > > bit_size(offset) > > does give the value of 64 Ok, good. > I reckon I had an error in my debug code, I was truncating the output format, > that explains why I'm chasing a gambuzino with this point. Or the OMPI bug

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Ricardo Reis
On Tue, 15 May 2012, Jeff Squyres wrote: On May 15, 2012, at 2:19 PM, Ricardo Reis wrote: INTEGER(kind=MPI_OFFSET_KIND) :: offset MPI_OFFSET_KIND is insuficient to represent my offset... Is it not a 64 bit integer for your compiler? I'm still interested in the answer to this question. p

Re: [OMPI users] mpirun error with " The MPI_Comm_f2c() function was called before MPI_INIT was invoked."

2012-05-15 Thread Jeff Squyres
More specifically -- are you calling any MPI function before calling MPI_Init? (particularly from Fortran) I.e., you may be seeing a side-effect of calling a fortran MPI function before calling MPI_Init. On May 13, 2012, at 9:53 AM, Ralph Castain wrote: > I believe the error message is pretty

Re: [OMPI users] Polling and Interrupt

2012-05-15 Thread Jeff Squyres
On May 14, 2012, at 11:33 AM, vaibhav dutt wrote: > Can anybody tell me how to enable the polling and interrupt/blocking > execution in > OpenMPI? Open MPI doesn't have a blocking mode; it always aggressively polls. -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: htt

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Jeff Squyres
Oops -- yes, I missed FILE_GET_POSITION_SHARED. I amended the patch: https://svn.open-mpi.org/trac/ompi/raw-attachment/ticket/3095/fortran-file-int-cast-fix2.diff MPI_Address is also a deprecated function exactly because it takes an fint -- MPI_Get_address replaced it, and uses an MPI_Aint. O

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Secretan Yves
Hy, The patch contains changes for mpi_file_get_view, mpi_file_get_position and mpi_file_get_size. I think mpi_file_get_position_shared is another candidate, and maybe mpi_address but this last one I am not sure In my first response, I did only list the function I use. Yves Secretan yves.secre

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Jeff Squyres
On May 15, 2012, at 2:19 PM, Ricardo Reis wrote: >>> INTEGER(kind=MPI_OFFSET_KIND) :: offset >>> >>> MPI_OFFSET_KIND is insuficient to represent my offset... >> >> Is it not a 64 bit integer for your compiler? I'm still interested in the answer to this question. >> There *is* a bug in OMPI at

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Ricardo Reis
On Tue, 15 May 2012, Jeff Squyres wrote: On May 15, 2012, at 10:53 AM, Ricardo Reis wrote: My problem is rather that INTEGER(kind=MPI_OFFSET_KIND) :: offset MPI_OFFSET_KIND is insuficient to represent my offset... Is it not a 64 bit integer for your compiler? There *is* a bug in OMPI at t

[OMPI users] mpicc link shouldn't add -ldl and -lhwloc

2012-05-15 Thread Orion Poplawski
See https://bugzilla.redhat.com/show_bug.cgi?id=814798 $ mpicc -showme:link -pthread -m64 -L/usr/lib64/openmpi/lib -lmpi -ldl -lhwloc -ldl and -lhwloc should not be listed. The user should only link against libraries that they are using directly, namely -lmpi, and they should explicitly add -

Re: [OMPI users] Open-MPI 1.6 searches for default hostfile in the wrong directory

2012-05-15 Thread Jeff Squyres
You are absolutely correct -- doh! I've fixed it on the trunk and submitted the fix for the v1.6 branch. Thanks! On May 14, 2012, at 9:14 AM, Götz Waschk wrote: > Dear Open-MPI developers, > > I have built my own package of openmpi 1.6 based on the RHEL6 package > on my SL6 test machine. My t

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Jeff Squyres
On May 15, 2012, at 10:53 AM, Ricardo Reis wrote: > My problem is rather that > > INTEGER(kind=MPI_OFFSET_KIND) :: offset > > MPI_OFFSET_KIND is insuficient to represent my offset... Is it not a 64 bit integer for your compiler? There *is* a bug in OMPI at the moment that we're casting the res

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Ricardo Reis
My problem is rather that INTEGER(kind=MPI_OFFSET_KIND) :: offset MPI_OFFSET_KIND is insuficient to represent my offset... best, Ricardo Reis 'Non Serviam' PhD/MSc Mechanical Engineering | Lic. Aerospace Engineering Computational Fluid Dynamics, High Performance Computing, Turbulen

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Jeff Squyres
On May 15, 2012, at 9:30 AM, Secretan Yves wrote: > If, by any bad luck, you use any of the following FORTRAN function > > MPI_FILE_GET_POSITION > MPI_FILE_GET_SIZE > MPI_FILE_GET_VIEW Ouch! Looking at the thread you cited, it looks like George never followed up on this. :-( I'll file a bu

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Secretan Yves
Hy, If, by any bad luck, you use any of the following FORTRAN function MPI_FILE_GET_POSITION MPI_FILE_GET_SIZE MPI_FILE_GET_VIEW MPI_TYPE_EXTENT they all are stiil overflowing (http://www.open-mpi.org/community/lists/devel/2010/12/8797.php) because they cast the correct result to MPI_Fint w

Re: [OMPI users] MPI-IO puzzlement

2012-05-15 Thread Ricardo Reis
Hi all The problem has been found. I'm trying to use MPI-IO to write the file with all processes taking part in the calculation writing their bit. Here lies the rub. Each process has to write a piece of DIM = 35709696 elements. Using 64 processes the ofsett is my_rank * dim and so...