flags are now stored in H5PartFile struct; added the 'throttle' as a runtime parameter instead of a compiler flag; merged H5Bench into the tests suite; compiles but need to perform regression tests

src/H5MultiBlock.c

@@ -906,29 +906,8 @@ _H5MultiBlock_read_data (
 	herr = _H5Block_select_hyperslab_for_reading ( f, dataset_id );
 	if ( herr < 0 ) return herr;
 
-#if H5PART_THROTTLE > 0
-	int ret;
-	int token = 1;
-	if (f->myproc == 0) {
-		_H5Part_print_info ("Using throttled reads with factor = %d",
-			H5PART_THROTTLE);
-	}
-	if (f->myproc % H5PART_THROTTLE > 0) {
-		_H5Part_print_debug_detail (
-			"[%d] Waiting on read token from %d",
-			f->myproc, f->myproc - 1);
-		// wait to receive token before continuing with read
-		ret = MPI_Recv(
-			&token, 1, MPI_INT,
-			f->myproc - 1, // receive from previous proc
-			f->myproc, // use this proc id as message tag
-			f->comm,
-			MPI_STATUS_IGNORE
-			);
-		if ( ret != MPI_SUCCESS ) return HANDLE_MPI_SENDRECV_ERR;
-	}
-	_H5Part_print_debug_detail ("[%d] Executing read", f->myproc);
-#endif
+	herr = _H5Part_start_throttle( f );
+	if (herr < 0) return herr;
 
 	herr = H5Dread ( 
 		dataset_id,
@@ -939,23 +918,8 @@ _H5MultiBlock_read_data (
 		*data );
 	if ( herr < 0 ) return HANDLE_H5D_READ_ERR ( field_name, f->timestep );
 
-#if H5PART_THROTTLE > 0
-	if (f->myproc % H5PART_THROTTLE < H5PART_THROTTLE - 1) {
-		// pass token to next proc 
-		if (f->myproc + 1 < f->nprocs) {
-			_H5Part_print_debug_detail (
-				"[%d] Passing read token to %d",
-				f->myproc, f->myproc + 1);
-			ret = MPI_Send(
-				&token, 1, MPI_INT,
-				f->myproc + 1, // send to next proc
-				f->myproc + 1, // use the id of the target as tag
-				f->comm
-				);
-		}
-		if ( ret != MPI_SUCCESS ) return HANDLE_MPI_SENDRECV_ERR;
-	}
-#endif
+	herr = _H5Part_end_throttle( f );
+	if (herr < 0) return herr;
 
 	herr = H5Dclose ( dataset_id );
 	if ( herr < 0 ) return HANDLE_H5D_CLOSE_ERR;
@@ -1056,30 +1020,9 @@ _H5MultiBlock_write_data (
 #endif
 	if ( dataset < 0 ) return HANDLE_H5D_CREATE_ERR ( name, f->timestep );
 
-#if H5PART_THROTTLE > 0
-	int ret;
-	int token = 1;
-	if (f->myproc == 0) {
-		_H5Part_print_info ("Using throttled writes with factor = %d",
-			H5PART_THROTTLE);
-	}
-	if (f->myproc % H5PART_THROTTLE > 0) {
-		_H5Part_print_debug_detail (
-			"[%d] Waiting on write token from %d",
-			f->myproc, f->myproc - 1);
-		// wait to receive token before continuing with write
-		ret = MPI_Recv(
-			&token, 1, MPI_INT,
-			f->myproc - 1, // receive from previous proc
-			f->myproc, // use this proc id as message tag
-			f->comm,
-			MPI_STATUS_IGNORE
-			);
-		if ( ret != MPI_SUCCESS ) return HANDLE_MPI_SENDRECV_ERR;
-	}
-	_H5Part_print_debug_detail ("[%d] Executing write", f->myproc);
-#endif
-		
+	herr = _H5Part_start_throttle( f );
+	if (herr < 0) return herr;
+
 	herr = H5Dwrite ( 
 		dataset,
 		type,
@@ -1089,23 +1032,8 @@ _H5MultiBlock_write_data (
 		data );
 	if ( herr < 0 ) return HANDLE_H5D_WRITE_ERR ( name, f->timestep );
 
-#if H5PART_THROTTLE > 0
-	if (f->myproc % H5PART_THROTTLE < H5PART_THROTTLE - 1) {
-		// pass token to next proc 
-		if (f->myproc + 1 < f->nprocs) {
-			_H5Part_print_debug_detail (
-				"[%d] Passing write token to %d",
-				f->myproc, f->myproc + 1);
-			ret = MPI_Send(
-				&token, 1, MPI_INT,
-				f->myproc + 1, // send to next proc
-				f->myproc + 1, // use the id of the target as tag
-				f->comm
-				);
-		}
-		if ( ret != MPI_SUCCESS ) return HANDLE_MPI_SENDRECV_ERR;
-	}
-#endif
+	herr = _H5Part_end_throttle( f );
+	if (herr < 0) return herr;
 
 	herr = H5Dclose ( dataset );
 	if ( herr < 0 ) return HANDLE_H5D_CLOSE_ERR;

src/H5MultiBlockErrors.h

@@ -22,12 +22,6 @@
 		H5PART_ERR_MPI, \
 		"Cannot create/commit/free strided vector MPI datatype." );
 
-#define HANDLE_MPI_SENDRECV_ERR \
-	 (*_err_handler) ( \
-		_H5Part_get_funcname(), \
-		H5PART_ERR_MPI, \
-		"Unable to perform point-to-point MPI send/receive." );
-
 #define HANDLE_MPI_INT64_ERR \
 	 (*_err_handler) ( \
 		_H5Part_get_funcname(), \

src/H5Part.c

@@ -147,6 +147,9 @@ _H5Part_open_file (
 	}
 	memset (f, 0, sizeof (H5PartFile));
 
+	f->flags = flags;
+	f->throttle = 0;
+
 	f->groupname_step = strdup ( H5PART_GROUPNAME_STEP );
 	if( f->groupname_step == NULL ) {
 		HANDLE_H5PART_NOMEM_ERR;
@@ -201,9 +204,13 @@ _H5Part_open_file (
 				goto error_cleanup;
 			}
 			if (flags & H5PART_VFD_MPIIO_IND) {
-				_H5Part_print_info ( "Using independent mode" );
+				if (f->myproc == 0) {
+					_H5Part_print_info ( "Using independent mode" );
+				}
 			} else {
-				_H5Part_print_info ( "Using collective mode" );
+				if (f->myproc == 0) {
+					_H5Part_print_info ( "Using collective mode" );
+				}
 				f->xfer_prop = H5Pcreate (H5P_DATASET_XFER);
 				if (f->xfer_prop < 0) {
 					HANDLE_H5P_CREATE_ERR;
@@ -762,6 +769,9 @@ _write_data (
 	if ( dataset_id < 0 )
 		return HANDLE_H5D_CREATE_ERR ( name, f->timestep );
 
+	herr = _H5Part_start_throttle( f );
+	if (herr < 0) return herr;
+
 	herr = H5Dwrite (
 		dataset_id,
 		type,
@@ -770,6 +780,9 @@ _write_data (
 		f->xfer_prop,
 		array );
 
+	herr = _H5Part_end_throttle( f );
+	if (herr < 0) return herr;
+
 	if ( herr < 0 ) return HANDLE_H5D_WRITE_ERR ( name, f->timestep );
 
 	herr = H5Dclose ( dataset_id );
@@ -2323,7 +2336,7 @@ h5part_int64_t
 H5PartHasView (
  	H5PartFile *f			/*!< [in]  Handle to open file */
 	) {
-	SET_FNAME ( "H5PartResetView" );
+	SET_FNAME ( "H5PartHasView" );
 
 	CHECK_FILEHANDLE( f );
 	CHECK_READONLY_MODE ( f );
@@ -2387,7 +2400,7 @@ _set_view (
 		return HANDLE_H5S_CREATE_SIMPLE_ERR ( total );
 
 	/* declare overall data size  but then will select a subset */
-	f->diskshape= H5Screate_simple ( 1, &total, &total );
+	f->diskshape = H5Screate_simple ( 1, &total, &total );
 	if ( f->diskshape < 0 )
 		return HANDLE_H5S_CREATE_SIMPLE_ERR ( total );
 
@@ -2598,17 +2611,9 @@ _read_data (
 	memspace_id = _get_memshape_for_reading ( f, dataset_id );
 	if ( memspace_id < 0 ) return (h5part_int64_t)memspace_id;
 
-#ifdef INDEPENDENT_IO
-	herr = H5Dread (
-		dataset_id,
-		type,
-		memspace_id,		/* shape/size of data in memory (the
-					   complement to disk hyperslab) */
-		space_id,		/* shape/size of data on disk 
-					   (get hyperslab if needed) */
-		H5P_DEFAULT,		/* ignore... its for parallel reads */
-		array );
-#else
+	herr = _H5Part_start_throttle( f );
+	if (herr < 0) return herr;
+
 	herr = H5Dread (
 		dataset_id,
 		type,
@@ -2618,7 +2623,9 @@ _read_data (
 					   (get hyperslab if needed) */
 		f->xfer_prop,		/* ignore... its for parallel reads */
 		array );
-#endif
+
+	herr = _H5Part_end_throttle( f );
+	if (herr < 0) return herr;
 
 	if ( herr < 0 ) return HANDLE_H5D_READ_ERR ( name, f->timestep );
 
@@ -2820,7 +2827,98 @@ H5PartReadParticleStep (
 	return H5PART_SUCCESS;
 }
 
-/****************** error handling ******************/
+/************ error handling and configuration ************/
+
+/*!
+  \ingroup h5part_errhandle
+
+  Set the `throttle` factor, which causes HDF5 write and read
+  calls to be issued in that number of batches.
+
+  This can prevent large cuncurrency parallel applications that
+  use independent writes from overwhelming the underlying
+  parallel file system.
+
+  Throttling only works with the H5PART_VFD_MPIPOSIX or
+  H5PART_VFD_MPIIO_IND drivers.
+
+  \return \c H5PART_SUCCESS
+*/
+h5part_int64_t
+H5PartSetThrottle (
+	H5PartFile *f,
+	int factor
+	) {
+
+	SET_FNAME( "H5PartSetThrottle" );
+	CHECK_FILEHANDLE ( f );
+
+	if (f->flags & H5PART_VFD_MPIIO_IND || f->flags & H5PART_VFD_MPIPOSIX) {
+		f->throttle = factor;
+	} else {
+		_H5Part_print_warn ("Throttling is not permitted in MPI-IO collective mode.");
+	}
+
+	return H5PART_SUCCESS;
+}
+
+h5part_int64_t
+_H5Part_start_throttle (
+	H5PartFile *f
+	) {
+
+	if (f->throttle > 0) {
+		int ret;
+		int token = 1;
+		if (f->myproc == 0) {
+			_H5Part_print_info ("Throttling with factor = %d",
+				f->throttle);
+		}
+		if (f->myproc % f->throttle > 0) {
+			_H5Part_print_debug_detail (
+				"[%d] throttle: waiting on token from %d",
+				f->myproc, f->myproc - 1);
+			// wait to receive token before continuing with read
+			ret = MPI_Recv(
+				&token, 1, MPI_INT,
+				f->myproc - 1, // receive from previous proc
+				f->myproc, // use this proc id as message tag
+				f->comm,
+				MPI_STATUS_IGNORE
+				);
+			if ( ret != MPI_SUCCESS ) return HANDLE_MPI_SENDRECV_ERR;
+		}
+		_H5Part_print_debug_detail ("[%d] throttle: received token", f->myproc);
+	}
+	return H5PART_SUCCESS;
+}
+
+h5part_int64_t
+_H5Part_end_throttle (
+	H5PartFile *f
+	) {
+
+	if (f->throttle > 0) {
+		int ret;
+		int token;
+		if (f->myproc % f->throttle < f->throttle - 1) {
+			// pass token to next proc 
+			if (f->myproc + 1 < f->nprocs) {
+				_H5Part_print_debug_detail (
+					"[%d] throttle: passing token to %d",
+					f->myproc, f->myproc + 1);
+				ret = MPI_Send(
+					&token, 1, MPI_INT,
+					f->myproc + 1, // send to next proc
+					f->myproc + 1, // use the id of the target as tag
+					f->comm
+					);
+			}
+			if ( ret != MPI_SUCCESS ) return HANDLE_MPI_SENDRECV_ERR;
+		}
+	}
+	return H5PART_SUCCESS;
+}
 
 /*!
   \ingroup h5part_errhandle

src/H5Part.h

@@ -349,11 +349,19 @@ H5PartReadFileAttrib (
 	void *value
 	);
 
+/*============ Error Reporting and Configuration =============*/
+
 h5part_int64_t
 H5PartSetVerbosityLevel (
 	const h5part_int64_t level
 	);
 
+h5part_int64_t
+H5PartSetThrottle (
+	H5PartFile *f,
+	int factor
+	);
+
 h5part_int64_t
 H5PartSetErrorHandler (
 	const h5part_error_handler handler

src/H5PartErrors.h

@@ -336,6 +336,12 @@
 		H5PART_ERR_MPI, \
 		"Cannot gather data." );
 
+#define HANDLE_MPI_SENDRECV_ERR \
+	 (*_err_handler) ( \
+		_H5Part_get_funcname(), \
+		H5PART_ERR_MPI, \
+		"Unable to perform point-to-point MPI send/receive." );
+
 #define HANDLE_MPI_COMM_SIZE_ERR \
 	 (*_err_handler) ( \
 		_H5Part_get_funcname(), \

src/H5PartPrivate.h

@@ -122,6 +122,16 @@ _H5Part_have_group (
 	const char *name
 	);
 
+h5part_int64_t
+_H5Part_start_throttle (
+	H5PartFile *f
+	);
+
+h5part_int64_t
+_H5Part_end_throttle (
+	H5PartFile *f
+	);
+
 void
 _H5Part_vprint_error (
 	const char *fmt,

src/H5PartTypes.h

@@ -19,7 +19,7 @@ __attribute__ ((format (printf, 3, 4)))
 #endif
  ;
 
-#ifndef PARALLEL_IO
+#ifndef MPI_INCLUDED
 typedef unsigned long		MPI_Comm;
 #endif
 
@@ -79,6 +79,9 @@ struct H5PartFile {
 	*/
 	MPI_Comm comm;
 
+	char flags;
+	int throttle;
+
 	struct H5BlockStruct *block;
 	h5part_int64_t (*close_block)(struct H5PartFile *f);
 

test/Makefile.am

@@ -53,6 +53,7 @@ EXTRA_DIST = 	Bench.c \
 		H5testFpar.F90 \
 		H5BlockParTestScalarField.c H5BlockParTestScalarFieldF.F90 \
 		H5BlockTestAttributes.c H5BlockTestAttributesF.F90 \
+                H5PartBench.c H5BlockBench.c \
 		$(bin_SCRIPTS)
 
 # Specific building instruction (What compilers to use...)
@@ -61,6 +62,12 @@ EXTRA_DIST = 	Bench.c \
 Bench: Bench.c
 	$(CC) $(CFLAGS) $(INC) -o $@ $< $(H5PLIB) -lH5Part $(LIBS)
 
+H5PartBench: H5PartBench.c
+	$(CC) $(CFLAGS) $(INC) -o $@ $< $(H5PLIB) -lpH5Part $(LIBS)
+
+H5BlockBench: H5BlockBench.c
+	$(CC) $(CFLAGS) $(INC) -o $@ $< $(H5PLIB) -lpH5Part $(LIBS)
+
 ###############################################################################
 
 H5PartTest: H5PartTest.o