[svn-r7860] Purpose:

    Bug fix

Description:
    Our previous "optimization" of metadata writing which only wrote metadata
from one process was abusing MPI-I/O and after some consultation with Rob Ross
and Rajeev Thakur, Albert & I have come up with a solution...

Solution:
    Instead of only writing from one process, issue a collective write
operation with all processes, for metadata writes.

Platforms tested:
    FreeBSD 4.9 (sleipnir)
    h5committest

4 files changed, 36 insertions, 0 deletions

diff --git a/src/H5F.c b/src/H5F.c
index c8d9bf2..574f512 100644
--- a/src/H5F.c
+++ b/src/H5F.c
@@ -237,6 +237,7 @@ H5F_init_interface(void)
     
     FUNC_ENTER_NOINIT(H5F_init_interface)
 
+#ifdef OLD_METADATA_WRITE
 #ifdef H5_HAVE_PARALLEL
     {
         /* Allow MPI buf-and-file-type optimizations? */
@@ -246,6 +247,7 @@ H5F_init_interface(void)
         }
     }
 #endif /* H5_HAVE_PARALLEL */
+#endif /* OLD_METADATA_WRITE */
 
     /*
      * Initialize the atom group for the file IDs. There are two groups:
diff --git a/src/H5FDmpio.c b/src/H5FDmpio.c
index 13d5ee7..2eda0ab 100644
--- a/src/H5FDmpio.c
+++ b/src/H5FDmpio.c
@@ -154,6 +154,7 @@ static int H5FD_mpio_Debug[256] =
           0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
 #endif
 
+#ifdef OLD_METADATA_WRITE
 /* Global var to allow elimination of redundant metadata writes
  * to be controlled by the value of an environment variable. */
 /* Use the elimination by default unless this is the Intel Red machine */
@@ -162,6 +163,7 @@ hbool_t	H5_mpi_1_metawrite_g = TRUE;
 #else
 hbool_t	H5_mpi_1_metawrite_g = FALSE;
 #endif
+#endif /* OLD_METADATA_WRITE */
 
 /* Interface initialization */
 #define INTERFACE_INIT	H5FD_mpio_init
@@ -1854,6 +1856,7 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
             if (MPI_SUCCESS!= (mpi_code=MPI_Barrier(file->comm)))
                 HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_code)
 
+#ifdef OLD_METADATA_WRITE
         /* Only p<round> will do the actual write if all procs in comm write same metadata */
         if (H5_mpi_1_metawrite_g) {
             if (file->mpi_rank != H5_PAR_META_WRITE) {
@@ -1867,6 +1870,22 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
                 HGOTO_DONE(SUCCEED) /* skip the actual write */
             }
         }
+#else /* OLD_METADATA_WRITE */
+        /* Remember that views are used */
+        use_view_this_time=TRUE;
+
+        /*
+         * Set the file view when we are using MPI derived types
+         */
+        /*OKAY: CAST DISCARDS CONST QUALIFIER*/
+        if (MPI_SUCCESS != (mpi_code=MPI_File_set_view(file->f, mpi_off, MPI_BYTE, MPI_BYTE, (char*)"native", file->info)))
+            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
+
+        /* When using types, use the address as the displacement for
+         * MPI_File_set_view and reset the address for the read to zero
+         */
+        mpi_off=0;
+#endif /* OLD_METADATA_WRITE */
     } /* end if */
 
     /* Write the data. */
@@ -1943,6 +1962,7 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
     file->eof = HADDR_UNDEF;
     
 done:
+#ifdef OLD_METADATA_WRITE
     /* Guard against getting into metadate broadcast in failure cases */
     if(ret_value!=FAIL) {
         /* if only p<round> writes, need to broadcast the ret_value to other processes */
@@ -1951,6 +1971,7 @@ done:
                 HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code)
         } /* end if */
     } /* end if */
+#endif /* OLD_METADATA_WRITE */
 
 #ifdef H5FDmpio_DEBUG
     if (H5FD_mpio_Debug[(int)'t'])
diff --git a/src/H5FDmpiposix.c b/src/H5FDmpiposix.c
index 2625208..647fb35 100644
--- a/src/H5FDmpiposix.c
+++ b/src/H5FDmpiposix.c
@@ -229,6 +229,7 @@ static const H5FD_class_t H5FD_mpiposix_g = {
     H5FD_FLMAP_SINGLE 				/*fl_map		*/
 };
 
+#ifdef OLD_METADATA_WRITE
 /* Global var to allow elimination of redundant metadata writes
  * to be controlled by the value of an environment variable. */
 /* Use the elimination by default unless this is the Intel Red machine */
@@ -237,6 +238,7 @@ hbool_t	H5_mpiposix_1_metawrite_g = TRUE;
 #else
 hbool_t	H5_mpiposix_1_metawrite_g = FALSE;
 #endif
+#endif /* OLD_METADATA_WRITE */
 
 /* Interface initialization */
 #define INTERFACE_INIT	H5FD_mpiposix_init
@@ -1309,7 +1311,9 @@ H5FD_mpiposix_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
                 HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_code)
 
         /* Only p<round> will do the actual write if all procs in comm write same metadata */
+#ifdef OLD_METADATA_WRITE
         if (H5_mpiposix_1_metawrite_g)
+#endif /* OLD_METADATA_WRITE */
             if (file->mpi_rank != H5_PAR_META_WRITE)
                 HGOTO_DONE(SUCCEED) /* skip the actual write */
     } /* end if */
@@ -1384,10 +1388,17 @@ done:
     /* Guard against getting into metadata broadcast in failure cases */
     else {
         /* if only p<round> writes, need to broadcast the ret_value to other processes */
+#ifdef OLD_METADATA_WRITE
         if ((type!=H5FD_MEM_DRAW) && H5_mpiposix_1_metawrite_g) {
             if (MPI_SUCCESS != (mpi_code= MPI_Bcast(&ret_value, sizeof(ret_value), MPI_BYTE, H5_PAR_META_WRITE, file->comm)))
                 HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code)
         } /* end if */
+#else /* OLD_METADATA_WRITE */
+        if (type!=H5FD_MEM_DRAW) {
+            if (MPI_SUCCESS != (mpi_code= MPI_Bcast(&ret_value, sizeof(ret_value), MPI_BYTE, H5_PAR_META_WRITE, file->comm)))
+                HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code)
+        } /* end if */
+#endif /* OLD_METADATA_WRITE */
     } /* end else */
 
     FUNC_LEAVE_NOAPI(ret_value)
diff --git a/src/H5Fpkg.h b/src/H5Fpkg.h
index ea11040..2ef630a 100644
--- a/src/H5Fpkg.h
+++ b/src/H5Fpkg.h
@@ -186,11 +186,13 @@ struct H5F_t {
     H5F_mtab_t		mtab;		/* File mount table		*/
 };
 
+#ifdef OLD_METADATA_WRITE
 #ifdef H5_HAVE_PARALLEL
 /* Whether a single process writes metadata */
 H5_DLLVAR  hbool_t H5_mpi_1_metawrite_g;
 H5_DLLVAR  hbool_t H5_mpiposix_1_metawrite_g;
 #endif /* H5_HAVE_PARALLEL */
+#endif /* OLD_METADATA_WRITE */
 
 /* Private functions, not part of the publicly documented API */
 #ifdef NOT_YET