Merge pull request #538 from wiredtiger/warm-cache-compress

Preload internal pages into file system cache with compression.

9 files changed, 127 insertions, 81 deletions

diff --git a/dist/stat_data.py b/dist/stat_data.py
index 2fc0373a6a3..be5efbe1caa 100644
--- a/dist/stat_data.py
+++ b/dist/stat_data.py
@@ -47,6 +47,7 @@ connection_stats = [
 	Stat('block_byte_read', 'bytes read by the block manager'),
 	Stat('block_byte_write', 'bytes written by the block manager'),
 	Stat('block_map_read', 'mapped blocks read by the block manager'),
+	Stat('block_preload', 'blocks pre-loaded by the block manager'),
 	Stat('block_read', 'blocks read by the block manager'),
 	Stat('block_write', 'blocks written by the block manager'),
 
diff --git a/src/block/block_mgr.c b/src/block/block_mgr.c
index f91a3079c3c..46bc223f70c 100644
--- a/src/block/block_mgr.c
+++ b/src/block/block_mgr.c
@@ -328,6 +328,7 @@ __bm_method_set(WT_BM *bm, int readonly)
 		    (WT_BM *, WT_SESSION_IMPL *, int, int *))__bm_readonly;
 		bm->free = (int (*)(WT_BM *,
 		    WT_SESSION_IMPL *, const uint8_t *, uint32_t))__bm_readonly;
+		bm->preload = __wt_bm_preload;
 		bm->read = __wt_bm_read;
 		bm->salvage_end = (int (*)
 		    (WT_BM *, WT_SESSION_IMPL *))__bm_readonly;
@@ -358,6 +359,7 @@ __bm_method_set(WT_BM *bm, int readonly)
 		bm->compact_page_skip = __bm_compact_page_skip;
 		bm->compact_skip = __bm_compact_skip;
 		bm->free = __bm_free;
+		bm->preload = __wt_bm_preload;
 		bm->read = __wt_bm_read;
 		bm->salvage_end = __bm_salvage_end;
 		bm->salvage_next = __bm_salvage_next;
diff --git a/src/block/block_read.c b/src/block/block_read.c
index bfb39dc567d..7e50435aac6 100644
--- a/src/block/block_read.c
+++ b/src/block/block_read.c
@@ -8,6 +8,52 @@
 #include "wt_internal.h"
 
 /*
+ * __wt_bm_preload --
+ *	Pre-load a page.
+ */
+int
+__wt_bm_preload(WT_BM *bm,
+    WT_SESSION_IMPL *session, const uint8_t *addr, uint32_t addr_size)
+{
+	WT_BLOCK *block;
+	WT_DECL_RET;
+	off_t offset;
+	uint32_t cksum, size;
+	int mapped;
+
+	WT_UNUSED(addr_size);
+	block = bm->block;
+
+	/* Crack the cookie. */
+	WT_RET(__wt_block_buffer_to_addr(block, addr, &offset, &size, &cksum));
+
+	/* Check for a mapped block. */
+	mapped = bm->map != NULL && offset + size <= (off_t)bm->maplen;
+	if (mapped)
+		WT_RET(__wt_mmap_preload(
+		    session, (uint8_t *)bm->map + offset, size));
+	else {
+#ifdef HAVE_POSIX_FADVISE
+		if ((ret = posix_fadvise(block->fh->fd,
+		    (off_t)offset, (off_t)size, POSIX_FADV_WILLNEED)) != 0)
+			WT_RET_MSG(
+			    session, ret, "%s: posix_fadvise", block->name);
+#else
+		WT_DECL_ITEM(tmp);
+		WT_RET(__wt_scr_alloc(session, size, &tmp));
+		ret = __wt_block_read_off(
+		    session, block, tmp, offset, size, cksum);
+		__wt_scr_free(&tmp);
+		WT_RET(ret);
+#endif
+	}
+
+	WT_CSTAT_INCR(session, block_preload);
+
+	return (0);
+}
+
+/*
  * __wt_bm_read --
  *	Map or read address cookie referenced block into a buffer.
  */
@@ -16,9 +62,9 @@ __wt_bm_read(WT_BM *bm, WT_SESSION_IMPL *session,
     WT_ITEM *buf, const uint8_t *addr, uint32_t addr_size)
 {
 	WT_BLOCK *block;
-	off_t offset;
-	uint32_t size, cksum;
 	int mapped;
+	off_t offset;
+	uint32_t cksum, size;
 
 	WT_UNUSED(addr_size);
 	block = bm->block;
@@ -48,7 +94,6 @@ __wt_bm_read(WT_BM *bm, WT_SESSION_IMPL *session,
 		buf->data = buf->mem;
 		buf->size = size;
 		F_SET(buf, WT_ITEM_MAPPED);
-
 		WT_RET(__wt_mmap_preload(session, buf->mem, buf->size));
 
 		WT_CSTAT_INCR(session, block_map_read);
diff --git a/src/btree/bt_handle.c b/src/btree/bt_handle.c
index 6ba3ec198f1..f8f58221c18 100644
--- a/src/btree/bt_handle.c
+++ b/src/btree/bt_handle.c
@@ -10,8 +10,8 @@
 static int __btree_conf(WT_SESSION_IMPL *, WT_CKPT *ckpt);
 static int __btree_get_last_recno(WT_SESSION_IMPL *);
 static int __btree_page_sizes(WT_SESSION_IMPL *);
+static int __btree_preload(WT_SESSION_IMPL *);
 static int __btree_tree_open_empty(WT_SESSION_IMPL *, int);
-static int __btree_warm_cache(WT_SESSION_IMPL *);
 
 static int pse1(WT_SESSION_IMPL *, const char *, uint32_t, uint32_t);
 static int pse2(WT_SESSION_IMPL *, const char *, uint32_t, uint32_t, uint32_t);
@@ -108,7 +108,7 @@ __wt_btree_open(WT_SESSION_IMPL *session, const char *op_cfg[])
 			    session, root_addr, root_addr_size));
 
 			/* Warm the cache, if possible. */
-			WT_ERR(__btree_warm_cache(session));
+			WT_ERR(__btree_preload(session));
 
 			/* Get the last record number in a column-store file. */
 			if (btree->type != BTREE_ROW)
@@ -490,39 +490,27 @@ __wt_btree_evictable(WT_SESSION_IMPL *session, int on)
 }
 
 /*
- * __btree_warm_cache --
- *	Pre-load internal pages from a checkpoint.
+ * __btree_preload --
+ *	Pre-load internal pages.
  */
 static int
-__btree_warm_cache(WT_SESSION_IMPL *session)
+__btree_preload(WT_SESSION_IMPL *session)
 {
-	WT_BTREE *btree;
 	WT_BM *bm;
-	WT_PAGE *page;
+	WT_BTREE *btree;
+	WT_REF *ref;
+	uint32_t addr_size, i;
+	const uint8_t *addr;
 
 	btree = S2BT(session);
 	bm = btree->bm;
 
-	if (bm->map == NULL || btree->compressor != NULL)
-		return (0);
-
-	/*
-	 * If the file is memory mapped, find the first leaf page.  Assuming
-	 * the file was created with a bulk load, the internal pages will be at
-	 * the end of the file, starting with the parent of the left-most
-	 * child, ending with the root.
-	 */
-	page = NULL;
-	WT_RET(__wt_tree_walk(session, &page, 0));
-	if (page == NULL)
-		return (WT_NOTFOUND);
-
-	if (page->parent->dsk < btree->root_page->dsk)
-		WT_RET(__wt_mmap_preload(
-		    session, page->parent->dsk,
-		    WT_PTRDIFF(btree->root_page->dsk, page->parent->dsk)));
-
-	return (__wt_page_release(session, page));
+	/* Pre-load the second-level internal pages. */
+	WT_REF_FOREACH(btree->root_page, ref, i) {
+		__wt_get_addr(btree->root_page, ref, &addr, &addr_size);
+		WT_RET(bm->preload(bm, session, addr, addr_size));
+	}
+	return (0);
 }
 
 /*
diff --git a/src/include/block.h b/src/include/block.h
index e41064d7134..277f6e871c4 100644
--- a/src/include/block.h
+++ b/src/include/block.h
@@ -151,6 +151,7 @@ struct __wt_bm {
 	    (WT_BM *, WT_SESSION_IMPL *, const uint8_t *, uint32_t, int *);
 	int (*compact_skip)(WT_BM *, WT_SESSION_IMPL *, int, int *);
 	int (*free)(WT_BM *, WT_SESSION_IMPL *, const uint8_t *, uint32_t);
+	int (*preload)(WT_BM *, WT_SESSION_IMPL *, const uint8_t *, uint32_t);
 	int (*read)
 	    (WT_BM *, WT_SESSION_IMPL *, WT_ITEM *, const uint8_t *, uint32_t);
 	int (*salvage_end)(WT_BM *, WT_SESSION_IMPL *);
diff --git a/src/include/extern.h b/src/include/extern.h
index 656173e6e06..244cbc22ad5 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -149,6 +149,10 @@ extern int __wt_desc_init(WT_SESSION_IMPL *session,
 extern void __wt_block_stat(WT_SESSION_IMPL *session,
     WT_BLOCK *block,
     WT_DSRC_STATS *stats);
+extern int __wt_bm_preload(WT_BM *bm,
+    WT_SESSION_IMPL *session,
+    const uint8_t *addr,
+    uint32_t addr_size);
 extern int __wt_bm_read(WT_BM *bm,
     WT_SESSION_IMPL *session,
     WT_ITEM *buf,
diff --git a/src/include/stat.h b/src/include/stat.h
index e6e598f5b19..83083874319 100644
--- a/src/include/stat.h
+++ b/src/include/stat.h
@@ -200,6 +200,7 @@ struct __wt_connection_stats {
 	WT_STATS block_byte_read;
 	WT_STATS block_byte_write;
 	WT_STATS block_map_read;
+	WT_STATS block_preload;
 	WT_STATS block_read;
 	WT_STATS block_write;
 	WT_STATS cache_bytes_dirty;
diff --git a/src/include/wiredtiger.in b/src/include/wiredtiger.in
index fbe181b0d5c..2005d208d44 100644
--- a/src/include/wiredtiger.in
+++ b/src/include/wiredtiger.in
@@ -2252,108 +2252,110 @@ extern int wiredtiger_extension_terminate(WT_CONNECTION *connection);
 #define	WT_STAT_CONN_BLOCK_BYTE_WRITE			2
 /*! mapped blocks read by the block manager */
 #define	WT_STAT_CONN_BLOCK_MAP_READ			3
+/*! blocks pre-loaded by the block manager */
+#define	WT_STAT_CONN_BLOCK_PRELOAD			4
 /*! blocks read by the block manager */
-#define	WT_STAT_CONN_BLOCK_READ				4
+#define	WT_STAT_CONN_BLOCK_READ				5
 /*! blocks written by the block manager */
-#define	WT_STAT_CONN_BLOCK_WRITE			5
+#define	WT_STAT_CONN_BLOCK_WRITE			6
 /*! cache: tracked dirty bytes in the cache */
-#define	WT_STAT_CONN_CACHE_BYTES_DIRTY			6
+#define	WT_STAT_CONN_CACHE_BYTES_DIRTY			7
 /*! cache: bytes currently in the cache */
-#define	WT_STAT_CONN_CACHE_BYTES_INUSE			7
+#define	WT_STAT_CONN_CACHE_BYTES_INUSE			8
 /*! cache: maximum bytes configured */
-#define	WT_STAT_CONN_CACHE_BYTES_MAX			8
+#define	WT_STAT_CONN_CACHE_BYTES_MAX			9
 /*! cache: bytes read into cache */
-#define	WT_STAT_CONN_CACHE_BYTES_READ			9
+#define	WT_STAT_CONN_CACHE_BYTES_READ			10
 /*! cache: bytes written from cache */
-#define	WT_STAT_CONN_CACHE_BYTES_WRITE			10
+#define	WT_STAT_CONN_CACHE_BYTES_WRITE			11
 /*! cache: checkpoint blocked page eviction */
-#define	WT_STAT_CONN_CACHE_EVICTION_CHECKPOINT		11
+#define	WT_STAT_CONN_CACHE_EVICTION_CHECKPOINT		12
 /*! cache: unmodified pages evicted */
-#define	WT_STAT_CONN_CACHE_EVICTION_CLEAN		12
+#define	WT_STAT_CONN_CACHE_EVICTION_CLEAN		13
 /*! cache: modified pages evicted */
-#define	WT_STAT_CONN_CACHE_EVICTION_DIRTY		13
+#define	WT_STAT_CONN_CACHE_EVICTION_DIRTY		14
 /*! cache: pages selected for eviction unable to be evicted */
-#define	WT_STAT_CONN_CACHE_EVICTION_FAIL		14
+#define	WT_STAT_CONN_CACHE_EVICTION_FAIL		15
 /*! cache: hazard pointer blocked page eviction */
-#define	WT_STAT_CONN_CACHE_EVICTION_HAZARD		15
+#define	WT_STAT_CONN_CACHE_EVICTION_HAZARD		16
 /*! cache: internal pages evicted */
-#define	WT_STAT_CONN_CACHE_EVICTION_INTERNAL		16
+#define	WT_STAT_CONN_CACHE_EVICTION_INTERNAL		17
 /*! cache: internal page merge operations completed */
-#define	WT_STAT_CONN_CACHE_EVICTION_MERGE		17
+#define	WT_STAT_CONN_CACHE_EVICTION_MERGE		18
 /*! cache: internal page merge attempts that could not complete */
-#define	WT_STAT_CONN_CACHE_EVICTION_MERGE_FAIL		18
+#define	WT_STAT_CONN_CACHE_EVICTION_MERGE_FAIL		19
 /*! cache: internal levels merged */
-#define	WT_STAT_CONN_CACHE_EVICTION_MERGE_LEVELS	19
+#define	WT_STAT_CONN_CACHE_EVICTION_MERGE_LEVELS	20
 /*! cache: eviction server unable to reach eviction goal */
-#define	WT_STAT_CONN_CACHE_EVICTION_SLOW		20
+#define	WT_STAT_CONN_CACHE_EVICTION_SLOW		21
 /*! cache: pages walked for eviction */
-#define	WT_STAT_CONN_CACHE_EVICTION_WALK		21
+#define	WT_STAT_CONN_CACHE_EVICTION_WALK		22
 /*! cache: tracked dirty pages in the cache */
-#define	WT_STAT_CONN_CACHE_PAGES_DIRTY			22
+#define	WT_STAT_CONN_CACHE_PAGES_DIRTY			23
 /*! cache: pages currently held in the cache */
-#define	WT_STAT_CONN_CACHE_PAGES_INUSE			23
+#define	WT_STAT_CONN_CACHE_PAGES_INUSE			24
 /*! cache: pages read into cache */
-#define	WT_STAT_CONN_CACHE_READ				24
+#define	WT_STAT_CONN_CACHE_READ				25
 /*! cache: pages written from cache */
-#define	WT_STAT_CONN_CACHE_WRITE			25
+#define	WT_STAT_CONN_CACHE_WRITE			26
 /*! pthread mutex condition wait calls */
-#define	WT_STAT_CONN_COND_WAIT				26
+#define	WT_STAT_CONN_COND_WAIT				27
 /*! cursor creation */
-#define	WT_STAT_CONN_CURSOR_CREATE			27
+#define	WT_STAT_CONN_CURSOR_CREATE			28
 /*! Btree cursor insert calls */
-#define	WT_STAT_CONN_CURSOR_INSERT			28
+#define	WT_STAT_CONN_CURSOR_INSERT			29
 /*! Btree cursor next calls */
-#define	WT_STAT_CONN_CURSOR_NEXT			29
+#define	WT_STAT_CONN_CURSOR_NEXT			30
 /*! Btree cursor prev calls */
-#define	WT_STAT_CONN_CURSOR_PREV			30
+#define	WT_STAT_CONN_CURSOR_PREV			31
 /*! Btree cursor remove calls */
-#define	WT_STAT_CONN_CURSOR_REMOVE			31
+#define	WT_STAT_CONN_CURSOR_REMOVE			32
 /*! Btree cursor reset calls */
-#define	WT_STAT_CONN_CURSOR_RESET			32
+#define	WT_STAT_CONN_CURSOR_RESET			33
 /*! Btree cursor search calls */
-#define	WT_STAT_CONN_CURSOR_SEARCH			33
+#define	WT_STAT_CONN_CURSOR_SEARCH			34
 /*! Btree cursor search near calls */
-#define	WT_STAT_CONN_CURSOR_SEARCH_NEAR			34
+#define	WT_STAT_CONN_CURSOR_SEARCH_NEAR			35
 /*! Btree cursor update calls */
-#define	WT_STAT_CONN_CURSOR_UPDATE			35
+#define	WT_STAT_CONN_CURSOR_UPDATE			36
 /*! files currently open */
-#define	WT_STAT_CONN_FILE_OPEN				36
+#define	WT_STAT_CONN_FILE_OPEN				37
 /*! rows merged in an LSM tree */
-#define	WT_STAT_CONN_LSM_ROWS_MERGED			37
+#define	WT_STAT_CONN_LSM_ROWS_MERGED			38
 /*! total heap memory allocations */
-#define	WT_STAT_CONN_MEMORY_ALLOCATION			38
+#define	WT_STAT_CONN_MEMORY_ALLOCATION			39
 /*! total heap memory frees */
-#define	WT_STAT_CONN_MEMORY_FREE			39
+#define	WT_STAT_CONN_MEMORY_FREE			40
 /*! total heap memory re-allocations */
-#define	WT_STAT_CONN_MEMORY_GROW			40
+#define	WT_STAT_CONN_MEMORY_GROW			41
 /*! total read I/Os */
-#define	WT_STAT_CONN_READ_IO				41
+#define	WT_STAT_CONN_READ_IO				42
 /*! page reconciliation calls */
-#define	WT_STAT_CONN_REC_PAGES				42
+#define	WT_STAT_CONN_REC_PAGES				43
 /*! page reconciliation calls for eviction */
-#define	WT_STAT_CONN_REC_PAGES_EVICTION			43
+#define	WT_STAT_CONN_REC_PAGES_EVICTION			44
 /*! reconciliation failed because an update could not be included */
-#define	WT_STAT_CONN_REC_SKIPPED_UPDATE			44
+#define	WT_STAT_CONN_REC_SKIPPED_UPDATE			45
 /*! pthread mutex shared lock read-lock calls */
-#define	WT_STAT_CONN_RWLOCK_READ			45
+#define	WT_STAT_CONN_RWLOCK_READ			46
 /*! pthread mutex shared lock write-lock calls */
-#define	WT_STAT_CONN_RWLOCK_WRITE			46
+#define	WT_STAT_CONN_RWLOCK_WRITE			47
 /*! open cursor count */
-#define	WT_STAT_CONN_SESSION_CURSOR_OPEN		47
+#define	WT_STAT_CONN_SESSION_CURSOR_OPEN		48
 /*! ancient transactions */
-#define	WT_STAT_CONN_TXN_ANCIENT			48
+#define	WT_STAT_CONN_TXN_ANCIENT			49
 /*! transactions */
-#define	WT_STAT_CONN_TXN_BEGIN				49
+#define	WT_STAT_CONN_TXN_BEGIN				50
 /*! transaction checkpoints */
-#define	WT_STAT_CONN_TXN_CHECKPOINT			50
+#define	WT_STAT_CONN_TXN_CHECKPOINT			51
 /*! transactions committed */
-#define	WT_STAT_CONN_TXN_COMMIT				51
+#define	WT_STAT_CONN_TXN_COMMIT				52
 /*! transaction failures due to cache overflow */
-#define	WT_STAT_CONN_TXN_FAIL_CACHE			52
+#define	WT_STAT_CONN_TXN_FAIL_CACHE			53
 /*! transactions rolled-back */
-#define	WT_STAT_CONN_TXN_ROLLBACK			53
+#define	WT_STAT_CONN_TXN_ROLLBACK			54
 /*! total write I/Os */
-#define	WT_STAT_CONN_WRITE_IO				54
+#define	WT_STAT_CONN_WRITE_IO				55
 
 /*!
  * @}
diff --git a/src/support/stat.c b/src/support/stat.c
index d62ec8dfbaf..a3ba49155ed 100644
--- a/src/support/stat.c
+++ b/src/support/stat.c
@@ -296,6 +296,7 @@ __wt_stat_init_connection_stats(WT_CONNECTION_STATS *stats)
 	stats->block_byte_read.desc = "bytes read by the block manager";
 	stats->block_byte_write.desc = "bytes written by the block manager";
 	stats->block_map_read.desc = "mapped blocks read by the block manager";
+	stats->block_preload.desc = "blocks pre-loaded by the block manager";
 	stats->block_read.desc = "blocks read by the block manager";
 	stats->block_write.desc = "blocks written by the block manager";
 	stats->cache_bytes_dirty.desc =
@@ -373,6 +374,7 @@ __wt_stat_clear_connection_stats(void *stats_arg)
 	stats->block_byte_read.v = 0;
 	stats->block_byte_write.v = 0;
 	stats->block_map_read.v = 0;
+	stats->block_preload.v = 0;
 	stats->block_read.v = 0;
 	stats->block_write.v = 0;
 	stats->cache_bytes_dirty.v = 0;