[svn-r9178] Purpose:

Fix parallel bug reported by Thomas Guignon, in which different
processes became confused as to whether they were doing collective
or individual I/O.


Description:
When one process had a point selection, and another didn't, the
first concluded that the I/O was independant, while the second
presumed that it was collective.  A hang resulted.


Solution:
Get all processes involved in an I/O to compare notes.  If all
agree that the I/O is collective, they proceed with same.  If
any think the I/O should be independant, all use independant
I/O.

Note that this is an interrim fix -- the correct solution is to
support collective I/O on point selections.  This will take a
while.


Platforms tested:
copper

h5committested

Eirene (serial and parallel)

In the parallel test on Eirene, I encountered a bug in h5repacktst.
However the problem vanished on recompile.  Since I couldn't reproduce
it elsewhere, I went ahead with the checkin.

Given my druthers, I would have liked to study the code more
carefully before this check-in.  However, there is some time pressure.

The new code implementing the consensus check must not be executed
unless MPI is initialized, and there is a communicator associated
with the file.  I think my guards against this case are adequate,
but if we run into a hang or an illegal instruction error, this
change should be suspect.


Misc. update:

1 files changed, 119 insertions, 0 deletions

diff --git a/src/H5S.c b/src/H5S.c
index 3ad61f8..ccc023c 100644
--- a/src/H5S.c
+++ b/src/H5S.c
@@ -33,6 +33,12 @@ static herr_t H5S_set_extent_simple (H5S_t *space, unsigned rank,
     const hsize_t *dims, const hsize_t *max);
 static htri_t H5S_is_simple(const H5S_t *sdim);
 
+#ifdef H5_HAVE_PARALLEL
+htri_t H5S_get_collective_io_consensus(const H5F_t *file,
+                                       const htri_t local_opinion,
+                                       const unsigned flags);
+#endif /* H5_HAVE_PARALLEL */
+
 /* Interface initialization */
 #define INTERFACE_INIT	H5S_init_interface
 static int		interface_initialize_g = 0;
@@ -1376,6 +1382,103 @@ done:
 
 
 /*-------------------------------------------------------------------------
+ * Function:	H5S_get_collective_io_consensus
+ *
+ * Purpose:	Compare notes with all other processes involved in this I/O
+ *		and see if all are go for collective I/O.  
+ *
+ *		If all are, return TRUE.  
+ *
+ *		If any process can't manage collective I/O, then collective
+ *		I/O is impossible, and we return FALSE.
+ *
+ *		If the flags indicate that collective I/O is impossible,
+ *		skip the interprocess communication and just return FALSE.
+ *
+ *		In any error is detected, return FAIL.
+ *		
+ * Return:	Success:	TRUE or FALSE
+ *
+ *		Failure:	FAIL
+ *
+ * Programmer:	JRM -- 8/30/04
+ *
+ * Modifications:
+ *
+ * 	None.
+ *
+ *-------------------------------------------------------------------------
+ */
+
+#ifdef H5_HAVE_PARALLEL
+htri_t
+H5S_get_collective_io_consensus(const H5F_t *file,
+                                const htri_t local_opinion,
+                                const unsigned flags)
+{
+    htri_t	ret_value = FAIL;	/* will update if successful */
+    MPI_Comm	comm;
+    int		int_local_opinion;
+    int		consensus;
+    int		mpi_result;
+    
+    FUNC_ENTER_NOAPI(H5S_get_collective_io_consensus, NULL);
+
+    HDassert ( ( local_opinion == TRUE ) || ( local_opinion == FALSE ) );
+
+    /* Don't do the interprocess communication unless the Parallel I/O 
+     * conversion flag is set -- there may not be other processes to 
+     * talk to.
+     */
+    if ( ! ( flags & flags&H5S_CONV_PAR_IO_POSSIBLE ) ) {
+
+	HGOTO_DONE(FALSE);
+    }
+
+    comm = H5F_mpi_get_comm(file);
+
+    if ( comm == MPI_COMM_NULL )
+        HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, \
+                    "can't retrieve MPI communicator")
+
+    if ( local_opinion == TRUE ) {
+
+        int_local_opinion = 1;
+
+    } else {
+
+        int_local_opinion = 0;
+    }
+
+    mpi_result = MPI_Allreduce((void *)(&int_local_opinion),
+                               (void *)(&consensus),
+                               1,
+                               MPI_INT,
+                               MPI_LAND,
+                               comm);
+
+    if ( mpi_result != MPI_SUCCESS ) 
+        HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_result)
+
+    if ( consensus ) {
+
+        ret_value = TRUE;
+
+    } else {
+
+        ret_value = FALSE;
+    }
+
+done:
+
+    FUNC_LEAVE_NOAPI(ret_value);
+
+} /* H5S_get_collective_io_consensus() */
+
+#endif /* H5_HAVE_PARALLEL */
+
+
+/*-------------------------------------------------------------------------
  * Function:	H5S_find
  *
  * Purpose:	Given two data spaces (MEM_SPACE and FILE_SPACE) this
@@ -1403,6 +1506,10 @@ done:
  *	along with other data whose scope is the conversion path (like path
  *	statistics).
  *
+ *	John Mainzer, 8/30/04
+ *	Modified code to check with all other processes that have the 
+ *	file open before OKing collective I/O.
+ *
  *-------------------------------------------------------------------------
  */
 H5S_conv_t *
@@ -1459,6 +1566,12 @@ const H5O_layout_t *layout
             if(opt==FAIL)
                 HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE, NULL, "invalid check for direct IO dataspace ");
 
+            opt = H5S_get_collective_io_consensus(file, opt, flags);
+
+            if ( opt == FAIL )
+                HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE, NULL, \
+                            "check for collective I/O consensus failed.");
+
             /* Check if we can use the optimized parallel I/O routines */
             if(opt==TRUE) {
                 /* Set the pointers to the MPI-specific routines */
@@ -1501,6 +1614,12 @@ const H5O_layout_t *layout
     if(opt==FAIL)
         HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE, NULL, "invalid check for direct IO dataspace ");
 
+    opt = H5S_get_collective_io_consensus(file, opt, flags);
+
+    if ( opt == FAIL )
+        HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE, NULL, \
+                    "check for collective I/O consensus failed.");
+
     /* Check if we can use the optimized parallel I/O routines */
     if(opt==TRUE) {
         /* Set the pointers to the MPI-specific routines */