[svn-r2990] Purpose:

    Bug fix (feature, kind of)
Description:
    The library used to hang if a collective dataset read/write request
    does not have the same number of eventual MPIO request.  Part of the
    reason is that H5FD_read/H5FD_write immediately returns succeess
    if it sees the request size is 0.  This caused problem since other
    processes with I/O to do would be hanging by waiting for the early
    returned process(es).
Solution:
H5FD.c: disable the early return code in parallel mode.  Make it
    go on even with "nothing" to transfer.
H5D.c: the optimized MPIO xfer routines can handle collect calls correctly
    when the condition is right (e.g., no conversion).  When the COLLECTIVE
    request cannot be handled correctly without the risk of hanging, the
    COLLECTIVE is changed to INDEPENDENT calls for the eventual MPIO
    calls.
Platforms tested:
    IRIX64 parallel (-64, n32), IRIX64 -64 sequential, Linux sequential.

1 files changed, 16 insertions, 0 deletions

diff --git a/src/H5FD.c b/src/H5FD.c
index aea02c2..8435963 100644
--- a/src/H5FD.c
+++ b/src/H5FD.c
@@ -1852,6 +1852,10 @@ H5FDread(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr, hsize_t siz
  *              Wednesday, August  4, 1999
  *
  * Modifications:
+ *	Albert Cheng, 2000-11-21
+ *	Disable the code that does early return when size==0 for
+ *	Parallel mode since a collective call would require the process
+ *	to continue on with "nothing" to transfer.
  *
  *-------------------------------------------------------------------------
  */
@@ -1865,8 +1869,12 @@ H5FD_read(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr, hsize_t si
 	   (H5P_DATA_XFER==H5P_get_class(dxpl_id) || H5I_object(dxpl_id)));
     assert(buf);
 
+#ifndef H5_HAVE_PARALLEL
+    /* Do not return early for Parallel mode since the I/O could be a */
+    /* collective transfer. */
     /* The no-op case */
     if (0==size) HRETURN(SUCCEED);
+#endif
 
     /* Check if this information is in the metadata accumulator */
     if((file->feature_flags&H5FD_FEAT_ACCUMULATE_METADATA) &&
@@ -2001,6 +2009,10 @@ H5FDwrite(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr, hsize_t si
  *              Wednesday, August  4, 1999
  *
  * Modifications:
+ *	Albert Cheng, 2000-11-21
+ *	Disable the code that does early return when size==0 for
+ *	Parallel mode since a collective call would require the process
+ *	to continue on with "nothing" to transfer.
  *
  *-------------------------------------------------------------------------
  */
@@ -2017,8 +2029,12 @@ H5FD_write(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr, hsize_t s
 	   (H5P_DATA_XFER==H5P_get_class(dxpl_id) && H5I_object(dxpl_id)));
     assert(buf);
     
+#ifndef H5_HAVE_PARALLEL
+    /* Do not return early for Parallel mode since the I/O could be a */
+    /* collective transfer. */
     /* The no-op case */
     if (0==size) HRETURN(SUCCEED);
+#endif
 
     /* Check for accumulating metadata */
     if((file->feature_flags&H5FD_FEAT_ACCUMULATE_METADATA) && type!=H5FD_MEM_DRAW) {