I'm looking at your code, and I'm not actually an expert in the MPI IO sutff... but do you have a race condition in the file close+delete and the open with EXCL?
I'm asking because I don't know offhand if the the file close+delete is supposed to be collective and not return until the file is guaranteed to be deleted, as visible from all MPI processes, or not. If this guarantee is not provided, then perhaps a barrier between the close+delete and the next file_open should be sufficient to avoid the race...? On Jan 15, 2014, at 7:26 PM, Ronald Cohen <rhco...@lbl.gov> wrote: > Update: I reconfigured with enable_io_romio=yes, and this time -- mostly -- > the test using Sample_mpio.c passes. Oddly the very first time I tried I > got errors: > > % mpirun -np 2 sampleio > Proc 1: hostname=Ron-Cohen-MBP.local > Testing simple C MPIO program with 2 processes accessing file ./mpitest.data > (Filename can be specified via program argument) > Proc 0: hostname=Ron-Cohen-MBP.local > Proc 1: read data[0:1] got 0, expect 1 > Proc 1: read data[0:2] got 0, expect 2 > Proc 1: read data[0:3] got 0, expect 3 > Proc 1: read data[0:4] got 0, expect 4 > Proc 1: read data[0:5] got 0, expect 5 > Proc 1: read data[0:6] got 0, expect 6 > Proc 1: read data[0:7] got 0, expect 7 > Proc 1: read data[0:8] got 0, expect 8 > Proc 1: read data[0:9] got 0, expect 9 > Proc 1: read data[1:0] got 0, expect 10 > Proc 1: read data[1:1] got 0, expect 11 > Proc 1: read data[1:2] got 0, expect 12 > Proc 1: read data[1:3] got 0, expect 13 > Proc 1: read data[1:4] got 0, expect 14 > Proc 1: read data[1:5] got 0, expect 15 > Proc 1: read data[1:6] got 0, expect 16 > Proc 1: read data[1:7] got 0, expect 17 > Proc 1: read data[1:8] got 0, expect 18 > Proc 1: read data[1:9] got 0, expect 19 > -------------------------------------------------------------------------- > MPI_ABORT was invoked on rank 1 in communicator MPI_COMM_WORLD > with errorcode 1. > > But when I reran the same mpirun command, the test was successful. And > deleting the executable and recompiling and then again running the same > mpirun command, the test was successful. Can someone explain that? > > > > > On Wed, Jan 15, 2014 at 1:16 PM, Ronald Cohen <rhco...@lbl.gov> wrote: > Aha. I guess I didn't know what the io-romio option does. If you look at > my config.log you will see my configure line included --disable-io-romio. > Guess I should change --disable to --enable. > > You seem to imply that the nightly build is stable enough that I should > probably switch to that rather than 1.7.4rc1. Am I reading between the > lines correctly? > > > > On Wed, Jan 15, 2014 at 10:56 AM, Ralph Castain <r...@open-mpi.org> wrote: > Oh, a word of caution on those config params - you might need to check to > ensure I don't disable romio in them. I don't normally build it as I don't > use it. Since that is what you are trying to use, just change the "no" to > "yes" (or delete that line altogether) and it will build. > > > > On Wed, Jan 15, 2014 at 10:53 AM, Ralph Castain <r...@open-mpi.org> wrote: > You can find my configure options in the OMPI distribution at > contrib/platform/intel/bend/mac. You are welcome to use them - just configure > --with-platform=intel/bend/mac > > I work on the developer's trunk, of course, but also run the head of the > 1.7.4 branch (essentially the nightly tarball) on a fairly regular basis. > > As for the opal_bitmap test: it wouldn't surprise me if that one was stale. I > can check on it later tonight, but I'd suspect that the test is bad as we use > that class in the code base and haven't seen an issue. > > > > On Wed, Jan 15, 2014 at 10:49 AM, Ronald Cohen <rhco...@lbl.gov> wrote: > Ralph, > > I just sent out another post with the c file attached. > > If you can get that to work, and even if you can't can you tell me what > configure options you use, and what version of open-mpi? Thanks. > > Ron > > > On Wed, Jan 15, 2014 at 10:36 AM, Ralph Castain <r...@open-mpi.org> wrote: > BTW: could you send me your sample test code? > > > On Wed, Jan 15, 2014 at 10:34 AM, Ralph Castain <r...@open-mpi.org> wrote: > I regularly build on Mavericks and run without problem, though I haven't > tried a parallel IO app. I'll give yours a try later, when I get back to my > Mac. > > > > On Wed, Jan 15, 2014 at 10:04 AM, Ronald Cohen <rhco...@lbl.gov> wrote: > I have been struggling trying to get a usable build of openmpi on Mac OSX > Mavericks (10.9.1). I can get openmpi to configure and build without error, > but have problems after that which depend on the openmpi version. > > With 1.6.5, make check fails the opal_datatype_test, ddt_test, and ddt_raw > tests. The various atomic_* tests pass. See checklogs_1.6.5, attached as > a .gz file. > > Following suggestions from openmpi discussions I tried openmpi version > 1.7.4rc1. In this case make check indicates all tests passed. But when I > proceeded to try to build a parallel code (parallel HDF5) it failed. > Following an email exchange with the HDF5 support people, they suggested I > try to compile and run the attached bit of simple code Sample_mpio.c (which > they supplied) which does not use any HDF5, but just attempts a parallel > write to a file and parallel read. That test failed when requesting more > than 1 processor -- which they say indicates a failure of the openmpi > installation. The error message was: > > MPI_INIT: argc 1 > MPI_INIT: argc 1 > Testing simple C MPIO program with 2 processes accessing file ./mpitest.data > (Filename can be specified via program argument) > Proc 0: hostname=Ron-Cohen-MBP.local > Proc 1: hostname=Ron-Cohen-MBP.local > MPI_BARRIER[0]: comm MPI_COMM_WORLD > MPI_BARRIER[1]: comm MPI_COMM_WORLD > Proc 0: MPI_File_open with MPI_MODE_EXCL failed (MPI_ERR_FILE: invalid file) > MPI_ABORT[0]: comm MPI_COMM_WORLD errorcode 1 > MPI_BCAST[1]: buffer 7fff5a483048 count 1 datatype MPI_INT root 0 comm > MPI_COMM_WORLD > > I then went back to my openmpi directories and tried running some of the > individual tests in the test and examples directories. In particular in > test/class I found one test that seem to not be run as part of make check > which failed, even with one processor; this is opal_bitmap. Not sure if this > is because 1.7.4rc1 is incomplete, or there is something wrong with the > installation, or maybe a 32 vs 64 bit thing? The error message is > > mpirun detected that one or more processes exited with non-zero status, thus > causing the job to be terminated. The first process to do so was: > > Process name: [[48805,1],0] > Exit code: 255 > > Any suggestions? > > More generally has anyone out there gotten an openmpi build on Mavericks to > work with sufficient success that they can get the attached Sample_mpio.c (or > better yet, parallel HDF5) to build? > > Details: Running Mac OS X 10.9.1 on a mid-2009 Macbook pro with 4 GB memory; > tried openmpi 1.6.5 and 1.7.4rc1. Built openmpi against the stock gcc that > comes with XCode 5.0.2, and gfortran 4.9.0. > > Files attached: config.log.gz, openmpialllog.gz (output of running ompi_info > --all), checklog2.gz (output of make.check in top openmpi directory). > > I am not attaching logs of make and install since those seem to have been > successful, but can generate those if that would be helpful. > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users > > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users > > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
