WT-2670 Add access_pattern_hint configuration for tables (#3155)

8 files changed, 121 insertions, 59 deletions

diff --git a/dist/api_config.py b/dist/api_config.py
index 1069a7037ce..d83a632321e 100644
--- a/dist/api_config.py
+++ b/dist/api_config.py
@@ -198,7 +198,8 @@ def get_default(c):
         return 'false'
     elif c.default == 'true':
         return 'true'
-    elif t == 'string' and c.default == 'none':
+    elif t == 'string' and c.default == 'none' and \
+        not c.flags.get('choices', []):
         return ''
     elif t == 'category':
         return '(%s)' % (','.join('%s=%s' % (subc.name, get_default(subc))
diff --git a/dist/api_data.py b/dist/api_data.py
index aa76ff45d1a..2d4371923e1 100644
--- a/dist/api_data.py
+++ b/dist/api_data.py
@@ -120,6 +120,14 @@ lsm_config = [
 
 # Per-file configuration
 file_config = format_meta + [
+    Config('access_pattern_hint', 'none', r'''
+        It is recommended that workloads that consist primarily of
+        updates and/or point queries specify \c random.  Workloads that
+        do many cursor scans through large ranges of data specify
+        \c sequential and other workloads specify \c none.  The
+        option leads to an advisory call to an appropriate operating
+        system API where available''',
+        choices=['none', 'random', 'sequential']),
     Config('block_allocation', 'best', r'''
         configure block allocation. Permitted values are \c "first" or
         \c "best"; the \c "first" configuration uses a first-available
diff --git a/src/block/block_open.c b/src/block/block_open.c
index 4d15942709a..07ceb4c8159 100644
--- a/src/block/block_open.c
+++ b/src/block/block_open.c
@@ -201,6 +201,12 @@ __wt_block_open(WT_SESSION_IMPL *session,
 	 * "direct_io=checkpoint" configures direct I/O for readonly data files.
 	 */
 	flags = 0;
+	WT_ERR(__wt_config_gets(session, cfg, "access_pattern_hint", &cval));
+	if (WT_STRING_MATCH("random", cval.str, cval.len))
+		LF_SET(WT_FS_OPEN_ACCESS_RAND);
+	else if (WT_STRING_MATCH("sequential", cval.str, cval.len))
+		LF_SET(WT_FS_OPEN_ACCESS_SEQ);
+
 	if (readonly && FLD_ISSET(conn->direct_io, WT_DIRECT_IO_CHECKPOINT))
 		LF_SET(WT_FS_OPEN_DIRECTIO);
 	if (!readonly && FLD_ISSET(conn->direct_io, WT_DIRECT_IO_DATA))
diff --git a/src/config/config_def.c b/src/config/config_def.c
index b3b900f8c42..9d886cbf0bd 100644
--- a/src/config/config_def.c
+++ b/src/config/config_def.c
@@ -224,6 +224,9 @@ static const WT_CONFIG_CHECK
 };
 
 static const WT_CONFIG_CHECK confchk_WT_SESSION_create[] = {
+	{ "access_pattern_hint", "string",
+	    NULL, "choices=[\"none\",\"random\",\"sequential\"]",
+	    NULL, 0 },
 	{ "allocation_size", "int",
 	    NULL, "min=512B,max=128MB",
 	    NULL, 0 },
@@ -401,6 +404,9 @@ static const WT_CONFIG_CHECK confchk_colgroup_meta[] = {
 };
 
 static const WT_CONFIG_CHECK confchk_file_config[] = {
+	{ "access_pattern_hint", "string",
+	    NULL, "choices=[\"none\",\"random\",\"sequential\"]",
+	    NULL, 0 },
 	{ "allocation_size", "int",
 	    NULL, "min=512B,max=128MB",
 	    NULL, 0 },
@@ -459,6 +465,9 @@ static const WT_CONFIG_CHECK confchk_file_config[] = {
 };
 
 static const WT_CONFIG_CHECK confchk_file_meta[] = {
+	{ "access_pattern_hint", "string",
+	    NULL, "choices=[\"none\",\"random\",\"sequential\"]",
+	    NULL, 0 },
 	{ "allocation_size", "int",
 	    NULL, "min=512B,max=128MB",
 	    NULL, 0 },
@@ -537,6 +546,9 @@ static const WT_CONFIG_CHECK confchk_index_meta[] = {
 };
 
 static const WT_CONFIG_CHECK confchk_lsm_meta[] = {
+	{ "access_pattern_hint", "string",
+	    NULL, "choices=[\"none\",\"random\",\"sequential\"]",
+	    NULL, 0 },
 	{ "allocation_size", "int",
 	    NULL, "min=512B,max=128MB",
 	    NULL, 0 },
@@ -1075,12 +1087,13 @@ static const WT_CONFIG_ENTRY config_entries[] = {
 	  confchk_WT_SESSION_compact, 1
 	},
 	{ "WT_SESSION.create",
-	  "allocation_size=4KB,app_metadata=,block_allocation=best,"
-	  "block_compressor=,cache_resident=false,checksum=uncompressed,"
-	  "colgroups=,collator=,columns=,dictionary=0,encryption=(keyid=,"
-	  "name=),exclusive=false,extractor=,format=btree,huffman_key=,"
-	  "huffman_value=,ignore_in_memory_cache_size=false,immutable=false"
-	  ",internal_item_max=0,internal_key_max=0,"
+	  "access_pattern_hint=none,allocation_size=4KB,app_metadata=,"
+	  "block_allocation=best,block_compressor=,cache_resident=false,"
+	  "checksum=uncompressed,colgroups=,collator=,columns=,dictionary=0"
+	  ",encryption=(keyid=,name=),exclusive=false,extractor=,"
+	  "format=btree,huffman_key=,huffman_value=,"
+	  "ignore_in_memory_cache_size=false,immutable=false,"
+	  "internal_item_max=0,internal_key_max=0,"
 	  "internal_key_truncate=true,internal_page_max=4KB,key_format=u,"
 	  "key_gap=10,leaf_item_max=0,leaf_key_max=0,leaf_page_max=32KB,"
 	  "leaf_value_max=0,log=(enabled=true),lsm=(auto_throttle=true,"
@@ -1090,7 +1103,7 @@ static const WT_CONFIG_ENTRY config_entries[] = {
 	  "os_cache_dirty_max=0,os_cache_max=0,prefix_compression=false,"
 	  "prefix_compression_min=4,source=,split_deepen_min_child=0,"
 	  "split_deepen_per_child=0,split_pct=75,type=file,value_format=u",
-	  confchk_WT_SESSION_create, 41
+	  confchk_WT_SESSION_create, 42
 	},
 	{ "WT_SESSION.drop",
 	  "checkpoint_wait=true,force=false,lock_wait=true,"
@@ -1172,10 +1185,26 @@ static const WT_CONFIG_ENTRY config_entries[] = {
 	  confchk_colgroup_meta, 5
 	},
 	{ "file.config",
-	  "allocation_size=4KB,app_metadata=,block_allocation=best,"
-	  "block_compressor=,cache_resident=false,checksum=uncompressed,"
-	  "collator=,columns=,dictionary=0,encryption=(keyid=,name=),"
-	  "format=btree,huffman_key=,huffman_value=,"
+	  "access_pattern_hint=none,allocation_size=4KB,app_metadata=,"
+	  "block_allocation=best,block_compressor=,cache_resident=false,"
+	  "checksum=uncompressed,collator=,columns=,dictionary=0,"
+	  "encryption=(keyid=,name=),format=btree,huffman_key=,"
+	  "huffman_value=,ignore_in_memory_cache_size=false,"
+	  "internal_item_max=0,internal_key_max=0,"
+	  "internal_key_truncate=true,internal_page_max=4KB,key_format=u,"
+	  "key_gap=10,leaf_item_max=0,leaf_key_max=0,leaf_page_max=32KB,"
+	  "leaf_value_max=0,log=(enabled=true),memory_page_max=5MB,"
+	  "os_cache_dirty_max=0,os_cache_max=0,prefix_compression=false,"
+	  "prefix_compression_min=4,split_deepen_min_child=0,"
+	  "split_deepen_per_child=0,split_pct=75,value_format=u",
+	  confchk_file_config, 35
+	},
+	{ "file.meta",
+	  "access_pattern_hint=none,allocation_size=4KB,app_metadata=,"
+	  "block_allocation=best,block_compressor=,cache_resident=false,"
+	  "checkpoint=,checkpoint_lsn=,checksum=uncompressed,collator=,"
+	  "columns=,dictionary=0,encryption=(keyid=,name=),format=btree,"
+	  "huffman_key=,huffman_value=,id=,"
 	  "ignore_in_memory_cache_size=false,internal_item_max=0,"
 	  "internal_key_max=0,internal_key_truncate=true,"
 	  "internal_page_max=4KB,key_format=u,key_gap=10,leaf_item_max=0,"
@@ -1183,24 +1212,8 @@ static const WT_CONFIG_ENTRY config_entries[] = {
 	  "log=(enabled=true),memory_page_max=5MB,os_cache_dirty_max=0,"
 	  "os_cache_max=0,prefix_compression=false,prefix_compression_min=4"
 	  ",split_deepen_min_child=0,split_deepen_per_child=0,split_pct=75,"
-	  "value_format=u",
-	  confchk_file_config, 34
-	},
-	{ "file.meta",
-	  "allocation_size=4KB,app_metadata=,block_allocation=best,"
-	  "block_compressor=,cache_resident=false,checkpoint=,"
-	  "checkpoint_lsn=,checksum=uncompressed,collator=,columns=,"
-	  "dictionary=0,encryption=(keyid=,name=),format=btree,huffman_key="
-	  ",huffman_value=,id=,ignore_in_memory_cache_size=false,"
-	  "internal_item_max=0,internal_key_max=0,"
-	  "internal_key_truncate=true,internal_page_max=4KB,key_format=u,"
-	  "key_gap=10,leaf_item_max=0,leaf_key_max=0,leaf_page_max=32KB,"
-	  "leaf_value_max=0,log=(enabled=true),memory_page_max=5MB,"
-	  "os_cache_dirty_max=0,os_cache_max=0,prefix_compression=false,"
-	  "prefix_compression_min=4,split_deepen_min_child=0,"
-	  "split_deepen_per_child=0,split_pct=75,value_format=u,"
-	  "version=(major=0,minor=0)",
-	  confchk_file_meta, 38
+	  "value_format=u,version=(major=0,minor=0)",
+	  confchk_file_meta, 39
 	},
 	{ "index.meta",
 	  "app_metadata=,collator=,columns=,extractor=,immutable=false,"
@@ -1208,23 +1221,23 @@ static const WT_CONFIG_ENTRY config_entries[] = {
 	  confchk_index_meta, 10
 	},
 	{ "lsm.meta",
-	  "allocation_size=4KB,app_metadata=,block_allocation=best,"
-	  "block_compressor=,cache_resident=false,checksum=uncompressed,"
-	  "chunks=,collator=,columns=,dictionary=0,encryption=(keyid=,"
-	  "name=),format=btree,huffman_key=,huffman_value=,"
-	  "ignore_in_memory_cache_size=false,internal_item_max=0,"
-	  "internal_key_max=0,internal_key_truncate=true,"
-	  "internal_page_max=4KB,key_format=u,key_gap=10,last=,"
-	  "leaf_item_max=0,leaf_key_max=0,leaf_page_max=32KB,"
-	  "leaf_value_max=0,log=(enabled=true),lsm=(auto_throttle=true,"
-	  "bloom=true,bloom_bit_count=16,bloom_config=,bloom_hash_count=8,"
-	  "bloom_oldest=false,chunk_count_limit=0,chunk_max=5GB,"
-	  "chunk_size=10MB,merge_max=15,merge_min=0),memory_page_max=5MB,"
-	  "old_chunks=,os_cache_dirty_max=0,os_cache_max=0,"
-	  "prefix_compression=false,prefix_compression_min=4,"
-	  "split_deepen_min_child=0,split_deepen_per_child=0,split_pct=75,"
-	  "value_format=u",
-	  confchk_lsm_meta, 38
+	  "access_pattern_hint=none,allocation_size=4KB,app_metadata=,"
+	  "block_allocation=best,block_compressor=,cache_resident=false,"
+	  "checksum=uncompressed,chunks=,collator=,columns=,dictionary=0,"
+	  "encryption=(keyid=,name=),format=btree,huffman_key=,"
+	  "huffman_value=,ignore_in_memory_cache_size=false,"
+	  "internal_item_max=0,internal_key_max=0,"
+	  "internal_key_truncate=true,internal_page_max=4KB,key_format=u,"
+	  "key_gap=10,last=,leaf_item_max=0,leaf_key_max=0,"
+	  "leaf_page_max=32KB,leaf_value_max=0,log=(enabled=true),"
+	  "lsm=(auto_throttle=true,bloom=true,bloom_bit_count=16,"
+	  "bloom_config=,bloom_hash_count=8,bloom_oldest=false,"
+	  "chunk_count_limit=0,chunk_max=5GB,chunk_size=10MB,merge_max=15,"
+	  "merge_min=0),memory_page_max=5MB,old_chunks=,"
+	  "os_cache_dirty_max=0,os_cache_max=0,prefix_compression=false,"
+	  "prefix_compression_min=4,split_deepen_min_child=0,"
+	  "split_deepen_per_child=0,split_pct=75,value_format=u",
+	  confchk_lsm_meta, 39
 	},
 	{ "table.meta",
 	  "app_metadata=,colgroups=,collator=,columns=,key_format=u,"
diff --git a/src/docs/upgrading.dox b/src/docs/upgrading.dox
index caa6b036892..c173bebdda6 100644
--- a/src/docs/upgrading.dox
+++ b/src/docs/upgrading.dox
@@ -17,6 +17,15 @@ how much work is done at the beginning of a checkpoint to make the critical
 section of checkpoints complete more quickly.
 </dd>
 
+<dt>Change to default fadvise setting for data files</dt>
+<dd>
+The default behavior for data files was to advise the file system to optimize
+for random access on POSIX and Windows platforms. The default is now to not
+advise about access patterns. There is a new \c access_pattern_hint
+configuration string available to WT_SESSION::create that can be used
+to configure the old default behavior.
+</dd>
+
 <dt>Checkpoint server created checkpoint names</dt>
 <dd>
 The ::wiredtiger_open checkpoint configuration no longer supports the
diff --git a/src/include/wiredtiger.in b/src/include/wiredtiger.in
index 37788096f39..2365135e08d 100644
--- a/src/include/wiredtiger.in
+++ b/src/include/wiredtiger.in
@@ -995,6 +995,13 @@ struct __wt_session {
 	 * \c "table:stock". For a description of URI formats
 	 * see @ref data_sources.
 	 * @configstart{WT_SESSION.create, see dist/api_data.py}
+	 * @config{access_pattern_hint, It is recommended that workloads that
+	 * consist primarily of updates and/or point queries specify \c random.
+	 * Workloads that do many cursor scans through large ranges of data
+	 * specify \c sequential and other workloads specify \c none.  The
+	 * option leads to an advisory call to an appropriate operating system
+	 * API where available., a string\, chosen from the following options:
+	 * \c "none"\, \c "random"\, \c "sequential"; default \c none.}
 	 * @config{allocation_size, the file unit allocation size\, in bytes\,
 	 * must a power-of-two; smaller values decrease the file space required
 	 * by overflow items\, and the default value of 4KB is a good choice
@@ -3734,21 +3741,25 @@ typedef enum {
 	WT_FS_OPEN_FILE_TYPE_REGULAR	/*!< open a regular file */
 } WT_FS_OPEN_FILE_TYPE;
 
+/*! WT_FILE_SYSTEM::open_file flags: random access pattern */
+#define	WT_FS_OPEN_ACCESS_RAND	0x001
+/*! WT_FILE_SYSTEM::open_file flags: sequential access pattern */
+#define	WT_FS_OPEN_ACCESS_SEQ	0x002
 /*! WT_FILE_SYSTEM::open_file flags: create if does not exist */
-#define	WT_FS_OPEN_CREATE	0x001
+#define	WT_FS_OPEN_CREATE	0x004
 /*! WT_FILE_SYSTEM::open_file flags: direct I/O requested */
-#define	WT_FS_OPEN_DIRECTIO	0x002
+#define	WT_FS_OPEN_DIRECTIO	0x008
 /*! WT_FILE_SYSTEM::open_file flags: file creation must be durable */
-#define	WT_FS_OPEN_DURABLE	0x004
+#define	WT_FS_OPEN_DURABLE	0x010
 /*!
  * WT_FILE_SYSTEM::open_file flags: return EBUSY if exclusive use not available
  */
-#define	WT_FS_OPEN_EXCLUSIVE	0x008
+#define	WT_FS_OPEN_EXCLUSIVE	0x020
 #ifndef DOXYGEN
-#define	WT_FS_OPEN_FIXED	0x010	/* Path not home relative (internal) */
+#define	WT_FS_OPEN_FIXED	0x040	/* Path not home relative (internal) */
 #endif
 /*! WT_FILE_SYSTEM::open_file flags: open is read-only */
-#define	WT_FS_OPEN_READONLY	0x020
+#define	WT_FS_OPEN_READONLY	0x080
 
 /*!
  * WT_FILE_SYSTEM::remove or WT_FILE_SYSTEM::rename flags: the remove or rename
diff --git a/src/os_posix/os_fs.c b/src/os_posix/os_fs.c
index 5f06892ce6e..26be0f049cc 100644
--- a/src/os_posix/os_fs.c
+++ b/src/os_posix/os_fs.c
@@ -575,7 +575,7 @@ __posix_open_file(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session,
 	WT_FILE_HANDLE_POSIX *pfh;
 	WT_SESSION_IMPL *session;
 	mode_t mode;
-	int f;
+	int advise_flag, f;
 
 	WT_UNUSED(file_system);
 
@@ -676,17 +676,25 @@ __posix_open_file(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session,
 
 #if defined(HAVE_POSIX_FADVISE)
 	/*
-	 * Disable read-ahead on trees: it slows down random read workloads.
+	 * If the user set an access pattern hint, call fadvise now.
 	 * Ignore fadvise when doing direct I/O, the kernel cache isn't
 	 * interesting.
 	 */
-	if (!pfh->direct_io && file_type == WT_FS_OPEN_FILE_TYPE_DATA) {
+	if (!pfh->direct_io && file_type == WT_FS_OPEN_FILE_TYPE_DATA &&
+	    LF_ISSET(WT_FS_OPEN_ACCESS_RAND | WT_FS_OPEN_ACCESS_SEQ)) {
+		advise_flag = 0;
+		if (LF_ISSET(WT_FS_OPEN_ACCESS_RAND))
+			advise_flag = POSIX_FADV_RANDOM;
+		if (LF_ISSET(WT_FS_OPEN_ACCESS_SEQ))
+			advise_flag = POSIX_FADV_SEQUENTIAL;
 		WT_SYSCALL(
-		    posix_fadvise(pfh->fd, 0, 0, POSIX_FADV_RANDOM), ret);
+		    posix_fadvise(pfh->fd, 0, 0, advise_flag), ret);
 		if (ret != 0)
 			WT_ERR_MSG(session, ret,
 			    "%s: handle-open: posix_fadvise", name);
 	}
+#else
+	WT_UNUSED(advise_flag);
 #endif
 
 directory_open:
diff --git a/src/os_win/os_fs.c b/src/os_win/os_fs.c
index 7ab7178114b..6c74f2f411f 100644
--- a/src/os_win/os_fs.c
+++ b/src/os_win/os_fs.c
@@ -521,10 +521,16 @@ __win_open_file(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session,
 	    FLD_ISSET(conn->txn_logsync, WT_LOG_DSYNC))
 		f |= FILE_FLAG_WRITE_THROUGH;
 
-	/* Disable read-ahead on trees: it slows down random read workloads. */
-	if (file_type == WT_FS_OPEN_FILE_TYPE_DATA)
+	/* If the user indicated a random workload, disable read-ahead. */
+	if (file_type == WT_FS_OPEN_FILE_TYPE_DATA &&
+	    LF_ISSET(WT_FS_OPEN_ACCESS_RAND))
 		f |= FILE_FLAG_RANDOM_ACCESS;
 
+	/* If the user indicated a sequential workload, set that. */
+	if (file_type == WT_FS_OPEN_FILE_TYPE_DATA &&
+	    LF_ISSET(WT_FS_OPEN_ACCESS_SEQ))
+		f |= FILE_FLAG_SEQUENTIAL_SCAN;
+
 	win_fh->filehandle = CreateFileW(name_wide->data, desired_access,
 	    FILE_SHARE_READ | FILE_SHARE_WRITE,
 	    NULL, dwCreationDisposition, f, NULL);