Support large single requests made by a process

benchmarks/C/parallel_run.sh

@@ -24,6 +24,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

benchmarks/FLASH-IO/parallel_run.sh

@@ -24,6 +24,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

configure.ac

@@ -139,13 +139,13 @@ AH_TEMPLATE([HDF5_VER_GE_1_10_4],       [Define if HDF5 version is at least 1.10
 AH_TEMPLATE([NETCDF_GE_4_5_0],          [Define if NetCDF version is at least 4.5.0])
 AH_TEMPLATE([PNC_MALLOC_TRACE],         [Define if to enable malloc tracing])
 AH_TEMPLATE([RELAX_COORD_BOUND],        [Define if relaxed coordinate check is enabled])
-AH_TEMPLATE([ENABLE_LARGE_SINGLE_REQ],  [Define if to enable large single MPI-IO request])
 AH_TEMPLATE([ENABLE_NULL_BYTE_HEADER_PADDING], [Define if to enable strict null-byte padding in file header])
 AH_TEMPLATE([ENABLE_BURST_BUFFER],      [Define if to enable burst buffer feature])
 AH_TEMPLATE([PNETCDF_PROFILING],        [Define if to enable PnetCDF internal performance profiling])
 AH_TEMPLATE([ENABLE_THREAD_SAFE],       [Define if to enable thread-safe capability])
 AH_TEMPLATE([ENABLE_REQ_AGGREGATION],   [Define if able to support request aggregation in nonblocking routines])
 dnl AH_TEMPLATE([HAVE_MPI_COUNT],       [Define if type MPI_Count is defined])
+AH_TEMPLATE([HAVE_MPI_LARGE_COUNT],     [Define if required MPI APIs have arguments of type MPI_Count])
 
 AH_TOP([#ifndef _CONFIG_H
 #define _CONFIG_H])
@@ -1315,6 +1315,26 @@ dnl                 MPI_Type_create_struct \
 dnl                 MPI_Type_create_resized \
 dnl                 MPI_Type_get_extent])
 
+dnl MPI_count was first introduced in MPI 3.0. Check MPI functions that make
+dnl use of MPI_Count.
+have_mpi_large_count_apis=yes
+AC_CHECK_FUNCS([MPI_Type_create_subarray_c \
+                MPI_Type_contiguous_c \
+                MPI_Type_create_hvector_c \
+                MPI_Type_create_struct_c \
+                MPI_Type_create_hindexed_c \
+                MPI_Type_vector_c \
+                MPI_Pack_c \
+                MPI_Unpack_c \
+                MPI_Type_get_true_extent_x \
+                MPI_Type_size_x], [], [have_mpi_large_count_apis=no])
+# If one of the above APIs is not available, have_mpi_large_count_apis will be
+# set to no
+UD_MSG_DEBUG([have_mpi_large_count_apis=$have_mpi_large_count_apis])
+if test "x$have_mpi_large_count_apis" = "xyes" ; then
+   AC_DEFINE(HAVE_MPI_LARGE_COUNT, 1)
+fi
+
 dnl Check presence of MPI COMBINERS. These are of type int.
 dnl These are introduced in MPI 2.0. As PnetCDF requires an MPI library that
 dnl supports MPI-IO and MPI-IO was first introduced in MPI 2.0, checking these
@@ -2256,21 +2276,6 @@ AC_SUBST(LATEX)
 AC_SUBST(DVIPDF)
 AM_CONDITIONAL([HAS_LATEX], [test "x$has_latex" = xyes])
 
-AC_ARG_ENABLE([large-single-req],
-    [AS_HELP_STRING([--enable-large-single-req],
-                    [Enable large (> 2 GiB) single request in individual MPI-IO
-                     calls. Note some MPI-IO libraries may not support this.
-                     @<:@default: disabled@:>@])],
-    [large_single_req=${enableval}], [large_single_req=no]
-)
-ENABLE_LARGE_SINGLE_REQ=0
-if test "x${large_single_req}" = xyes ; then
-    ENABLE_LARGE_SINGLE_REQ=1
-    AC_DEFINE(ENABLE_LARGE_SINGLE_REQ)
-fi
-AM_CONDITIONAL(ENABLE_LARGE_SINGLE_REQ, [test x$large_single_req = xyes])
-AC_SUBST(ENABLE_LARGE_SINGLE_REQ)
-
 AC_ARG_ENABLE([large-file-test],
     [AS_HELP_STRING([--enable-large-file-test],
                     [Enable testing for large (>4GB) file/variable I/O. Note
@@ -2722,10 +2727,6 @@ if test "x${enable_subfiling}" = xyes; then
    echo "\
               Subfiling support                           - enabled"
 fi
-if test "x${large_single_req}" = xyes ; then
-   echo "\
-              Allow large (> 2GiB) single MPI-IO requests - enabled"
-fi
 if test "x${thread_safe}" = xyes ; then
    echo "\
               Thread-safe capability                      - enabled"

examples/C/parallel_run.sh

@@ -27,6 +27,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

examples/CXX/parallel_run.sh

@@ -27,6 +27,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

examples/F77/parallel_run.sh

@@ -27,6 +27,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

examples/F90/parallel_run.sh

@@ -27,6 +27,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

examples/adios/parallel_run.sh

@@ -23,6 +23,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

examples/burst_buffer/parallel_run.sh

@@ -27,6 +27,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     # echo "---- exec=$i"
     for j in ${safe_modes} ; do

examples/tutorial/parallel_run.sh

@@ -27,6 +27,9 @@ else
    safe_modes="0"
 fi
 
+# prevent user environment setting of PNETCDF_HINTS to interfere
+unset PNETCDF_HINTS
+
 for i in ${check_PROGRAMS} ; do
     for j in ${safe_modes} ; do
         if test "$j" = 1 ; then # test only in safe mode

sneak_peek.md

@@ -14,14 +14,26 @@ This is essentially a placeholder for the next release note ...
     ```
     ncmpi_put_vara_float_all(ncid, varid, start, count, buf);
     ```
+  + PnetCDF now allows a single read/write request from a process of size
+    larger than 2 GiB. Such large requests are now passed down to the MP-IO
+    library. This change is because MPI 3.0 introduces the large count feature,
+    including MPI_Count data type, MPI_XXX_c and MPI_XXX_x APIs that use 8-byte
+    integer type to enable large MPI operations. As many MPI libraries today
+    have implemented this feature, PnetCDF can now take advantage of it to
+    support large single requests. Because of this change configure option
+    `--enable-large-single-req` is thus deprecated. See
+    See [PR #131](https://github.com/Parallel-NetCDF/PnetCDF/pull/131)
 
 * New optimization
   + none
 
 * New Limitations
-  + none
+  + Hint `nc_header_read_chunk_size` is limited to `NC_MAX_INT`. PnetCDF reads
+    file header in chunks. This hint customizes the chunk size.
 
 * Configure options
+  + `--enable-large-single-req` has been removed, as PnetCDF now allows a
+    single reqd/write request of size larger than 2 GiB.
   + `--disable-file-sync` is now deprecated. This configure option alone does
     not provide a sufficient data consistency. Users are suggested to call
     `ncmpi_sync` and `MPI_Barrier` to achieve a desired consistency.
@@ -146,6 +158,7 @@ This is essentially a placeholder for the next release note ...
     See [PR #79](https://github.com/Parallel-NetCDF/PnetCDF/pull/79).
 
 * Bug fixes
+  + Fix hint values that are actually used. See commit 41e8ef8.
   + Fix residual values of `v_align` and `r_align` when re-entering the define
     mode multiple times.
     See [PR #126](https://github.com/Parallel-NetCDF/PnetCDF/pull/126).
@@ -166,6 +179,9 @@ This is essentially a placeholder for the next release note ...
   + none
 
 * New test program
+  + test/largefile/large_header.c - test file header size larger than 2 GiB.
+  + test/largefile/large_reqs.c - test a single read/write request of size
+    larger than 2 GiB.
   + test/testcases/tst_redefine.c - test multiple entries of `ncmpi__enddef`
     [PR #126](https://github.com/Parallel-NetCDF/PnetCDF/pull/126).
   + test/testcases/tst_symlink.c - test `NC_CLOBBER` on a symbolic link.

src/binding/f77/pnetcdf.inc.in

@@ -37,7 +37,6 @@
       parameter (PNETCDF_SUBFILING                = @ENABLE_SUBFILING@)
       parameter (PNETCDF_RELAX_COORD_BOUND        = @RELAX_COORD_BOUND@)
       parameter (PNETCDF_DEBUG_MODE               = @PNETCDF_DEBUG@)
-      parameter (PNETCDF_LARGE_SINGLE_REQ         = @ENABLE_LARGE_SINGLE_REQ@)
       parameter (PNETCDF_NULL_BYTE_HEADER_PADDING = @ENABLE_NULL_BYTE_HEADER_PADDING@)
       parameter (PNETCDF_BYTE_SWAP_IN_PLACE       = @IN_PLACE_SWAP@) ! -1 means auto
       parameter (PNETCDF_BURST_BUFFERING          = @ENABLE_BURST_BUFFER@)

src/binding/f90/nfmpi_constants.fh.in

@@ -24,7 +24,6 @@
       PNETCDF_SUBFILING                = @ENABLE_SUBFILING@, &
       PNETCDF_RELAX_COORD_BOUND        = @RELAX_COORD_BOUND@, &
       PNETCDF_DEBUG_MODE               = @PNETCDF_DEBUG@, &
-      PNETCDF_LARGE_SINGLE_REQ         = @ENABLE_LARGE_SINGLE_REQ@, &
       PNETCDF_NULL_BYTE_HEADER_PADDING = @ENABLE_NULL_BYTE_HEADER_PADDING@, &
       PNETCDF_BYTE_SWAP_IN_PLACE       = @IN_PLACE_SWAP@, & ! -1 means auto
       PNETCDF_BURST_BUFFERING          = @ENABLE_BURST_BUFFER@, &

src/drivers/common/create_imaptype.c

@@ -27,7 +27,7 @@ int
 ncmpii_create_imaptype(int               ndims,
                        const MPI_Offset *count,   /* [ndims] */
                        const MPI_Offset *imap,    /* [ndims] */
-                       MPI_Datatype      ptype,   /* element type in buftype */
+                       MPI_Datatype      itype,   /* element type */
                        MPI_Datatype     *imaptype)/* out */
 {
     int dim, el_size, mpireturn;
@@ -55,7 +55,8 @@ ncmpii_create_imaptype(int               ndims,
     if (dim == -1) /* imap is a contiguous layout */
         return NC_NOERR;
 
-    MPI_Type_size(ptype, &el_size);
+    /* itype: element data type (MPI primitive type) */
+    MPI_Type_size(itype, &el_size);
 
     /* We have a true varm call, as imap gives non-contiguous layout.
      * User buffer will be packed (write case) or unpacked (read case)
@@ -66,17 +67,28 @@ ncmpii_create_imaptype(int               ndims,
      * dim is the first dimension (C order, eg. ZYX) that has
      * non-contiguous imap.
      */
-    if (imap_contig_blocklen != (int)imap_contig_blocklen)
-        DEBUG_RETURN_ERROR(NC_EINTOVERFLOW)
-    if (count[dim] != (int)count[dim] || imap[dim] != (int)imap[dim])
+    if (imap_contig_blocklen > NC_MAX_INT || count[dim] > NC_MAX_INT ||
+        imap[dim] > NC_MAX_INT) {
+#ifdef HAVE_MPI_LARGE_COUNT
+        mpireturn = MPI_Type_vector_c(count[dim], imap_contig_blocklen,
+                                      imap[dim], itype, imaptype);
+        if (mpireturn != MPI_SUCCESS) {
+            ncmpii_error_mpi2nc(mpireturn,"MPI_Type_vector_c");
+            DEBUG_RETURN_ERROR(NC_EMPI)
+        }
+#else
         DEBUG_RETURN_ERROR(NC_EINTOVERFLOW)
-
-    mpireturn = MPI_Type_vector((int)count[dim], (int)imap_contig_blocklen,
-                                (int)imap[dim], ptype, imaptype);
-    if (mpireturn != MPI_SUCCESS) {
-        ncmpii_error_mpi2nc(mpireturn,"MPI_Type_vector");
-        DEBUG_RETURN_ERROR(NC_EMPI)
+#endif
     }
+    else {
+        mpireturn = MPI_Type_vector((int)count[dim], (int)imap_contig_blocklen,
+                                    (int)imap[dim], itype, imaptype);
+        if (mpireturn != MPI_SUCCESS) {
+            ncmpii_error_mpi2nc(mpireturn,"MPI_Type_vector");
+            DEBUG_RETURN_ERROR(NC_EMPI)
+        }
+    }
+
     mpireturn = MPI_Type_commit(imaptype);
     if (mpireturn != MPI_SUCCESS) {
         ncmpii_error_mpi2nc(mpireturn,"MPI_Type_commit");
@@ -85,14 +97,26 @@ ncmpii_create_imaptype(int               ndims,
 
     for (dim--; dim>=0; dim--) {
         MPI_Datatype tmptype;
-        if (count[dim] != (int)count[dim])
-            DEBUG_RETURN_ERROR(NC_EINTOVERFLOW)
 
-        mpireturn = MPI_Type_create_hvector((int)count[dim], 1,
-                    imap[dim]*el_size, *imaptype, &tmptype);
-        if (mpireturn != MPI_SUCCESS) {
-            ncmpii_error_mpi2nc(mpireturn,"MPI_Type_create_hvector");
-            DEBUG_RETURN_ERROR(NC_EMPI)
+        if (count[dim] > NC_MAX_INT) {
+#ifdef HAVE_MPI_LARGE_COUNT
+            mpireturn = MPI_Type_create_hvector_c(count[dim], 1,
+                        imap[dim]*el_size, *imaptype, &tmptype);
+            if (mpireturn != MPI_SUCCESS) {
+                ncmpii_error_mpi2nc(mpireturn,"MPI_Type_create_hvector_c");
+                DEBUG_RETURN_ERROR(NC_EMPI)
+            }
+#else
+            DEBUG_RETURN_ERROR(NC_EINTOVERFLOW)
+#endif
+        }
+        else {
+            mpireturn = MPI_Type_create_hvector((int)count[dim], 1,
+                        imap[dim]*el_size, *imaptype, &tmptype);
+            if (mpireturn != MPI_SUCCESS) {
+                ncmpii_error_mpi2nc(mpireturn,"MPI_Type_create_hvector");
+                DEBUG_RETURN_ERROR(NC_EMPI)
+            }
         }
 
         mpireturn = MPI_Type_free(imaptype);

src/drivers/common/dtype_decode.c

@@ -71,22 +71,21 @@ dtype_filter(MPI_Datatype type)
 
     /* 4-byte integer types */
   {
-    MPI_Datatype int_4byte;
 #if (SIZEOF_LONG == 4)
     MPI_Datatype uint_4byte;
 #endif
 #if (SIZEOF_INT == 4)
-     int_4byte = MPI_INT;
+     MPI_Datatype int_4byte = MPI_INT;
 #if (SIZEOF_LONG == 4)
     uint_4byte = MPI_UNSIGNED;
 #endif
 #elif (SIZEOF_SHORT == 4)
-     int_4byte = MPI_SHORT;
+     MPI_Datatype int_4byte = MPI_SHORT;
 #if (SIZEOF_LONG == 4)
     uint_4byte = MPI_UNSIGNED_SHORT;
 #endif
 #else
-     int_4byte = MPI_DATATYPE_NULL; /* no 4-byte type? */
+     MPI_Datatype int_4byte = MPI_DATATYPE_NULL; /* no 4-byte type? */
 #if (SIZEOF_LONG == 4)
     uint_4byte = MPI_DATATYPE_NULL;
 #endif