mpi error from ncmpi_end_indep_data

[Wei-keng Liao]

Hi, Jim

Just a reminder, all PnetCDF nonblocking APIs can be called in both
collective and independent data modes. Only the wait APIs must be called
in its data mode, i.e. ncmpi_wait() in independent mode and ncmpi_wait_all()
in collective mode. So, you can safely remove the begin and end indep_data()
in PIO. In particular, ncmpi_end_indep_data() can be expensive and should
be avoided if possible.

Wei-keng

On Dec 30, 2014, at 1:12 PM, Jim Edwards wrote:

> Okay great - thank you!
> 
> On Tue, Dec 30, 2014 at 12:11 PM, Wei-keng Liao <wkliao at eecs.northwestern.edu> wrote:
> Hi, Jim
> 
> After I move the call to ncmpi_wait() to after ncmpi_end_indep_data()
> I can reproduce the error.
> 
> The problem has been fixed in the latest revision of PnetCDF.
> 
> 
> Wei-keng
> 
> On Dec 30, 2014, at 12:13 PM, Wei-keng Liao wrote:
> 
> > Hi, Jim
> >
> > I was not able to reproduce the error. Could you try the following program?
> > Also, could you add an error checking for the ncmpi APIs before ncmpi_end_indep_data.
> > I wonder if there is an error returned from one of those APIs.
> >
> > #include <stdio.h>
> > #include <stdlib.h>
> > #include <mpi.h>
> > #include <pnetcdf.h>
> >
> > #define NY 4
> > #define NX 10
> > #define NDIMS 2
> >
> > #define ERR \
> >    if (err != NC_NOERR) { \
> >        printf("Error at line=%d: %s\n", __LINE__, ncmpi_strerror(err)); \
> >    }
> >
> > int main(int argc, char** argv)
> > {
> >    int rank, nprocs, err;
> >    int ncid, varid, dimid[2], req, st;
> >    MPI_Offset start[2], count[2], stride[2];
> >    unsigned char buffer[NY][NX];
> >
> >    MPI_Init(&argc, &argv);
> >    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
> >    MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
> >
> >    err = ncmpi_create(MPI_COMM_WORLD, "testfile.nc", NC_CLOBBER|NC_64BIT_DATA,
> >                       MPI_INFO_NULL, &ncid);
> >    ERR
> >
> >    err = ncmpi_def_dim(ncid, "Y", NY,        &dimid[0]); ERR
> >    err = ncmpi_def_dim(ncid, "X", NX*nprocs, &dimid[1]); ERR
> >    err = ncmpi_def_var(ncid, "var", NC_UBYTE, NDIMS, dimid, &varid); ERR
> >    err = ncmpi_enddef(ncid); ERR
> >
> >     start[0] = 0;     start[1] = NX*rank;
> >     count[0] = NY/2;  count[1] = NX/2;
> >    stride[0] = 2;    stride[1] = 2;
> >    err = ncmpi_buffer_attach(ncid, NY*NX); ERR
> >
> >    err = ncmpi_begin_indep_data(ncid); ERR
> >    err = ncmpi_bput_vars_uchar(ncid, varid, start, count, stride,
> >                                &buffer[0][0], &req);
> >    ERR
> >    err = ncmpi_wait(ncid, 1, &req, &st); ERR
> >    err = ncmpi_end_indep_data(ncid); ERR
> >
> >    err = ncmpi_buffer_detach(ncid); ERR
> >    err = ncmpi_close(ncid); ERR
> >
> >    MPI_Finalize();
> >    return 0;
> > }
> >
> >
> > Wei-keng
> >
> > On Dec 30, 2014, at 10:31 AM, Jim Edwards wrote:
> >
> >> Hi Wei-keng,
> >>
> >> I have a code block that looks like:
> >>
> >>      ncmpi_begin_indep_data(file->fh);
> >>
> >>      usage = 0;
> >>
> >>      if(ios->io_rank==file->indep_rank){
> >>
> >> ierr = ncmpi_bput_vars_uchar(file->fh, varid, start, count, stride, op, &request);;
> >>
> >> pio_push_request(file, request);
> >>
> >> ierr = ncmpi_inq_buffer_usage(ncid, &usage);
> >>
> >> //   printf("%s %d %d\n",__FILE__,__LINE__,usage);
> >>
> >>      }
> >>
> >>
> >>      ncmpi_end_indep_data(file->fh);
> >>
> >>
> >> It's generating an error message on the ncmpi_end_indep_data call:
> >> MPI_FILE_WRITE_AT(83): File does not exist
> >>
> >> Regardless of the error message the write seems to work fine, any idea what's causing this?
> >>
> >> I'm using pnetcdf svn revision 1920 with intel 15 compiler and mpich.
> >>
> >> - Jim
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >> --
> >> Jim Edwards
> >>
> >> CESM Software Engineer
> >> National Center for Atmospheric Research
> >> Boulder, CO
> >
> 
> 
> 
> 
> -- 
> Jim Edwards
> 
> CESM Software Engineer
> National Center for Atmospheric Research
> Boulder, CO