[svn-r12311] Purpose:

Add pinned entry capability to cache.


Description:

For frequently accessed cache entries, the protect/unprotect overhead
is sometimes a bottleneck.


Solution:

Allow entries to be pinned in the cache.  Pinned entries can't be
evicted, but can be flushed or modified.


Platforms tested:

h5committested -- minus one small typo in test/cache.c whose fix was
                  tested on copper and heping only.


Misc. update:

1 files changed, 315 insertions, 0 deletions

diff --git a/src/H5ACpkg.h b/src/H5ACpkg.h
new file mode 100644
index 0000000..903eca3
--- /dev/null
+++ b/src/H5ACpkg.h
@@ -0,0 +1,315 @@
+/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
+ * Copyright by the Board of Trustees of the University of Illinois.         *
+ * All rights reserved.                                                      *
+ *                                                                           *
+ * This file is part of HDF5.  The full HDF5 copyright notice, including     *
+ * terms governing use, modification, and redistribution, is contained in    *
+ * the files COPYING and Copyright.html.  COPYING can be found at the root   *
+ * of the source code distribution tree; Copyright.html can be found at the  *
+ * root level of an installed copy of the electronic HDF5 document set and   *
+ * is linked from the top-level documents page.  It can also be found at     *
+ * http://hdf.ncsa.uiuc.edu/HDF5/doc/Copyright.html.  If you do not have     *
+ * access to either file, you may request a copy from hdfhelp@ncsa.uiuc.edu. *
+ * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
+
+/*
+ * Programmer: John Mainzer -- 4/19/06
+ *
+ * Purpose:     This file contains declarations which are normally visible
+ *              only within the H5AC package (just H5AC.c at present).
+ *
+ *		Source files outside the H5AC package should include
+ *		H5ACprivate.h instead.
+ *
+ *		The one exception to this rule is testpar/t_cache.c.  The 
+ *		test code is easier to write if it can look at H5AC_aux_t.
+ *		Indeed, this is the main reason why this file was created.
+ */
+
+#ifndef H5AC_PACKAGE
+#error "Do not include this file outside the H5AC package!"
+#endif
+
+#ifndef _H5ACpkg_H
+#define _H5ACpkg_H
+
+#define H5C_PACKAGE             /*suppress error about including H5Cpkg   */
+
+/* Get package's private header */
+#include "H5Cprivate.h"
+
+
+/* Get needed headers */
+#include "H5Cpkg.h"             /* Cache                                */
+#include "H5SLprivate.h"        /* Skip lists */
+
+#ifdef H5_HAVE_PARALLEL
+#include <mpi.h>
+#endif /* H5_HAVE_PARALLEL */
+
+
+#define H5AC_DEBUG_DIRTY_BYTES_CREATION	0
+
+/*-------------------------------------------------------------------------
+ *  It is a bit difficult to set ranges of allowable values on the 
+ *  dirty_bytes_threshold field of H5AC_aux_t.  The following are 
+ *  probably broader than they should be.  
+ *-------------------------------------------------------------------------
+ */
+
+#define H5AC__MIN_DIRTY_BYTES_THRESHOLD		(int32_t) \
+						(H5C__MIN_MAX_CACHE_SIZE / 2)
+#define H5AC__DEFAULT_DIRTY_BYTES_THRESHOLD	(256 * 1024)
+#define H5AC__MAX_DIRTY_BYTES_THRESHOLD   	(int32_t) \
+						(H5C__MAX_MAX_CACHE_SIZE / 4)
+
+/****************************************************************************
+ *
+ * structure H5AC_aux_t
+ *
+ * While H5AC has become a wrapper for the cache implemented in H5C.c, there
+ * are some features of the metadata cache that are specific to it, and which
+ * therefore do not belong in the more generic H5C cache code.
+ *
+ * In particular, there is the matter of synchronizing writes from the 
+ * metadata cache to disk in the PHDF5 case.
+ *
+ * Prior to this update, the presumption was that all metadata caches would 
+ * write the same data at the same time since all operations modifying 
+ * metadata must be performed collectively.  Given this assumption, it was 
+ * safe to allow only the writes from process 0 to actually make it to disk, 
+ * while metadata writes from all other processes were discarded.
+ *
+ * Unfortunately, this presumption is in error as operations that read 
+ * metadata need not be collective, but can change the location of dirty 
+ * entries in the metadata cache LRU lists.  This can result in the same 
+ * metadata write operation triggering writes from the metadata caches on 
+ * some processes, but not all (causing a hang), or in different sets of 
+ * entries being written from different caches (potentially resulting in 
+ * metadata corruption in the file).
+ *
+ * To deal with this issue, I decided to apply a paradigm shift to the way
+ * metadata is written to disk.
+ *
+ * With this set of changes, only the metadata cache on process 0 is able 
+ * to write metadata to disk, although metadata caches on all other 
+ * processes can read metadata from disk as before.
+ *
+ * To keep all the other caches from getting plugged up with dirty metadata,
+ * process 0 periodically broadcasts a list of entries that it has flushed
+ * since that last notice, and which are currently clean.  The other caches
+ * mark these entries as clean as well, which allows them to evict the 
+ * entries as needed.
+ *
+ * One obvious problem in this approach is synchronizing the broadcasts
+ * and receptions, as different caches may see different amounts of 
+ * activity.  
+ *
+ * The current solution is for the caches to track the number of bytes 
+ * of newly generated dirty metadata, and to broadcast and receive 
+ * whenever this value exceeds some user specified threshold.
+ *
+ * Maintaining this count is easy for all processes not on process 0 --
+ * all that is necessary is to add the size of the entry to the total 
+ * whenever there is an insertion, a rename of a previously clean entry,
+ * or whever a previously clean entry is marked dirty in an unprotect.
+ *
+ * On process 0, we have to be careful not to count dirty bytes twice.
+ * If an entry is marked dirty, flushed, and marked dirty again, all 
+ * within a single reporting period, it only th first marking should 
+ * be added to the dirty bytes generated tally, as that is all that 
+ * the other processes will see.
+ *
+ * At present, this structure exists to maintain the fields needed to
+ * implement the above scheme, and thus is only used in the parallel
+ * case.  However, other uses may arise in the future.
+ *
+ * Instance of this structure are associated with metadata caches via 
+ * the aux_ptr field of H5C_t (see H5Cpkg.h).  The H5AC code is 
+ * responsible for allocating, maintaining, and discarding instances
+ * of H5AC_aux_t. 
+ *
+ * The remainder of this header comments documents the individual fields
+ * of the structure.
+ *
+ *                                              JRM - 6/27/05
+ *
+ * magic:       Unsigned 32 bit integer always set to 
+ *		H5AC__H5AC_AUX_T_MAGIC.  This field is used to validate 
+ *		pointers to instances of H5AC_aux_t.
+ *
+ * mpi_comm:	MPI communicator associated with the file for which the
+ *		cache has been created.
+ *
+ * mpi_rank:	MPI rank of this process within mpi_comm.
+ *
+ * mpi_size:	Number of processes in mpi_comm.
+ *
+ * write_permitted:  Boolean flag used to control whether the cache
+ *		is permitted to write to file.  
+ *
+ * dirty_bytes_threshold: Integer field containing the dirty bytes 
+ *		generation threashold.  Whenever dirty byte creation 
+ *		exceeds this value, the metadata cache on process 0 
+ *		broadcasts a list of the entries it has flushed since
+ *		the last broadcast (or since the beginning of execution)
+ *		and which are currently clean (if they are still in the 
+ *		cache)
+ *
+ *		Similarly, metadata caches on processes other than process
+ *		0 will attempt to receive a list of clean entries whenever
+ *		the threshold is exceeded.
+ *
+ * dirty_bytes:  Integer field containing the number of bytes of dirty
+ *		metadata generated since the beginning of the computation, 
+ *		or (more typically) since the last clean entries list 
+ *		broadcast.  This field is reset to zero after each such
+ *		broadcast.
+ *
+ * dirty_bytes_propagations: This field only exists when the 
+ *		H5AC_DEBUG_DIRTY_BYTES_CREATION #define is TRUE.
+ *
+ *		It is used to track the number of times the cleaned list
+ *		has been propagated from process 0 to the other 
+ *		processes.
+ *
+ * unprotect_dirty_bytes:  This field only exists when the
+ *              H5AC_DEBUG_DIRTY_BYTES_CREATION #define is TRUE.
+ *
+ *		It is used to track the number of dirty bytes created
+ *		via unprotect operations since the last time the cleaned
+ *		list was propagated.
+ *
+ * unprotect_dirty_bytes_updates: This field only exists when the
+ *              H5AC_DEBUG_DIRTY_BYTES_CREATION #define is TRUE.
+ *
+ *		It is used to track the number of times dirty bytes have
+ *		been created via unprotect operations since the last time 
+ *		the cleaned list was propagated.
+ *
+ * insert_dirty_bytes:  This field only exists when the
+ *              H5AC_DEBUG_DIRTY_BYTES_CREATION #define is TRUE.
+ *
+ *		It is used to track the number of dirty bytes created
+ *		via insert operations since the last time the cleaned
+ *		list was propagated.
+ *
+ * insert_dirty_bytes_updates:  This field only exists when the
+ *              H5AC_DEBUG_DIRTY_BYTES_CREATION #define is TRUE.
+ *
+ *		It is used to track the number of times dirty bytes have
+ *		been created via insert operations since the last time 
+ *		the cleaned list was propagated.
+ *
+ * rename_dirty_bytes:  This field only exists when the
+ *              H5AC_DEBUG_DIRTY_BYTES_CREATION #define is TRUE.
+ *
+ *		It is used to track the number of dirty bytes created
+ *		via rename operations since the last time the cleaned
+ *		list was propagated.
+ *
+ * rename_dirty_bytes_updates:  This field only exists when the
+ *              H5AC_DEBUG_DIRTY_BYTES_CREATION #define is TRUE.
+ *
+ *		It is used to track the number of times dirty bytes have
+ *		been created via rename operations since the last time 
+ *		the cleaned list was propagated.
+ *
+ * d_slist_ptr:  Pointer to an instance of H5SL_t used to maintain a list
+ *		of entries that have been dirtied since the last time they
+ *		were listed in a clean entries broadcast.  This list is
+ *		only maintained by the metadata cache on process 0 -- it
+ *		it used to maintain a view of the dirty entries as seen
+ *		by the other caches, so as to keep the dirty bytes count
+ *		in synchronization with them.
+ *
+ *		Thus on process 0, the dirty_bytes count is incremented
+ *		only if either
+ *
+ *		1) an entry is inserted in the metadata cache, or
+ *
+ *		2) a previously clean entry is renamed, and it does not
+ *		   already appear in the dirty entry list, or
+ *
+ *		3) a previously clean entry is unprotected with the 
+ *		   dirtied flag set and the entry does not already appear 
+ *		   in the dirty entry list.
+ *
+ *		Entries are added to the dirty entry list whever they cause
+ *		the dirty bytes count to be increased.  They are removed 
+ *		when they appear in a clean entries broadcast.  Note that
+ *		renames must be reflected in the dirty entry list.
+ *
+ *		To reitterate, this field is only used on process 0 -- it 
+ *		should be NULL on all other processes.
+ *
+ * d_slist_len: Integer field containing the number of entries in the 
+ *		dirty entry list.  This field should always contain the 
+ *		value 0 on all processes other than process 0.  It exists
+ *		primarily for sanity checking.
+ *
+ * c_slist_ptr: Pointer to an instance of H5SL_t used to maintain a list 
+ *		of entries that were dirty, have been flushed
+ *		to disk since the last clean entries broadcast, and are
+ *		still clean.  Since only process 0 can write to disk, this
+ *		list only exists on process 0.
+ *
+ *		In essence, this slist is used to assemble the contents of
+ *		the next clean entries broadcast.  The list emptied after 
+ *		each broadcast.
+ *		
+ * c_slist_len: Integer field containing the number of entries in the clean
+ *		entries list (*c_slist_ptr).  This field should always 
+ *		contain the value 0 on all processes other than process 0.  
+ *		It exists primarily for sanity checking.
+ *
+ ****************************************************************************/
+
+#ifdef H5_HAVE_PARALLEL
+
+#define H5AC__H5AC_AUX_T_MAGIC        (unsigned)0x00D0A01
+
+typedef struct H5AC_aux_t
+{
+    uint32_t	magic;
+
+    MPI_Comm	mpi_comm;
+
+    int		mpi_rank;
+
+    int		mpi_size;
+
+    hbool_t	write_permitted;
+
+    int32_t	dirty_bytes_threshold;
+
+    int32_t	dirty_bytes;
+
+#if H5AC_DEBUG_DIRTY_BYTES_CREATION 
+
+    int32_t	dirty_bytes_propagations;
+
+    int32_t     unprotect_dirty_bytes;
+    int32_t     unprotect_dirty_bytes_updates;
+
+    int32_t     insert_dirty_bytes;
+    int32_t     insert_dirty_bytes_updates;
+
+    int32_t     rename_dirty_bytes;
+    int32_t     rename_dirty_bytes_updates;
+
+#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */
+
+    H5SL_t *	d_slist_ptr;
+
+    int32_t	d_slist_len;
+
+    H5SL_t *	c_slist_ptr;
+
+    int32_t	c_slist_len;
+
+} H5AC_aux_t; /* struct H5AC_aux_t */
+
+#endif /* H5_HAVE_PARALLEL */
+
+#endif /* _H5Cpkg_H */