branches/innodb+ rb://48

This patch is to improve recovery performance in InnoDB+.
It includes introduction of red-black tree for sorted insertion into
the flush_list and couple of other quirks. More can be found
at: https://svn.innodb.com/innobase/Recovery_Performance_Improvements

Reviewed by: Marko

buf/buf0flu.c

@@ -37,6 +37,142 @@ buf_flush_validate_low(void);
 		/* out: TRUE if ok */
 #endif /* UNIV_DEBUG || UNIV_BUF_DEBUG */
 
+/**********************************************************************
+Insert a block in the flush_rbt and returns a pointer to its
+predecessor or NULL if no predecessor. The ordering is maintained
+on the basis of the <oldest_modification, space, offset> key. */
+static
+buf_page_t*
+buf_flush_insert_in_flush_rbt(
+/*==========================*/
+				/* out: pointer to the predecessor or
+				NULL if no predecessor. */
+	buf_page_t*	bpage)	/* in: bpage to be inserted. */
+{
+	buf_page_t*		prev = NULL;
+	const ib_rbt_node_t*	c_node;
+	const ib_rbt_node_t*	p_node;
+
+	ut_ad(buf_pool_mutex_own());
+
+	/* Insert this buffer into the rbt. */
+	c_node = rbt_insert(buf_pool->flush_rbt, &bpage, &bpage);
+	ut_a(c_node != NULL);
+
+	/* Get the predecessor. */
+	p_node = rbt_prev(buf_pool->flush_rbt, c_node);
+
+	if (p_node != NULL) {
+		prev = *rbt_value(buf_page_t*, p_node);
+		ut_a(prev != NULL);
+	}
+
+	return(prev);
+}
+
+/*************************************************************
+Delete a bpage from the flush_rbt. */
+static
+void
+buf_flush_delete_from_flush_rbt(
+/*============================*/
+	buf_page_t*	bpage)	/* in: bpage to be removed. */
+{
+
+	ibool	ret = FALSE;
+
+	ut_ad(buf_pool_mutex_own());
+	ret = rbt_delete(buf_pool->flush_rbt, &bpage);
+	ut_ad(ret);
+}
+
+/*********************************************************************
+Compare two modified blocks in the buffer pool. The key for comparison
+is:
+key = <oldest_modification, space, offset>
+This comparison is used to maintian ordering of blocks in the
+buf_pool->flush_rbt.
+Note that for the purpose of flush_rbt, we only need to order blocks
+on the oldest_modification. The other two fields are used to uniquely
+identify the blocks. */
+static
+int
+buf_flush_block_cmp(
+/*================*/
+					/* out:
+					< 0 if b2 < b1,
+					0 if b2 == b1,
+					> 0 if b2 > b1 */
+	const void*	p1,		/* in: block1 */
+	const void*	p2)		/* in: block2 */
+{
+	int		ret;
+
+	ut_ad(p1 != NULL);
+	ut_ad(p2 != NULL);
+
+	const buf_page_t* b1 = *(const buf_page_t**) p1;
+	const buf_page_t* b2 = *(const buf_page_t**) p2;
+
+	ut_ad(b1 != NULL);
+	ut_ad(b2 != NULL);
+
+	ut_ad(b1->in_flush_list);
+	ut_ad(b2->in_flush_list);
+
+	if (b2->oldest_modification
+	    > b1->oldest_modification) {
+		return(1);
+	}
+
+	if (b2->oldest_modification
+	    < b1->oldest_modification) {
+		return(-1);
+	}
+
+	/* If oldest_modification is same then decide on the space. */
+	ret = (int)(b2->space - b1->space);
+
+	/* Or else decide ordering on the offset field. */
+	return(ret ? ret : (int)(b2->offset - b1->offset));
+}
+
+/************************************************************************
+Initialize the red-black tree to speed up insertions into the flush_list
+during recovery process. Should be called at the start of recovery
+process before any page has been read/written. */
+UNIV_INTERN
+void
+buf_flush_init_flush_rbt(void)
+/*==========================*/
+{
+	buf_pool_mutex_enter();
+
+	/* Create red black tree for speedy insertions in flush list. */
+	buf_pool->flush_rbt = rbt_create(sizeof(buf_page_t*),
+					 buf_flush_block_cmp);
+	buf_pool_mutex_exit();
+}
+
+/************************************************************************
+Frees up the red-black tree. */
+UNIV_INTERN
+void
+buf_flush_free_flush_rbt(void)
+/*==========================*/
+{
+	buf_pool_mutex_enter();
+
+#if defined UNIV_DEBUG || defined UNIV_BUF_DEBUG
+	ut_a(buf_flush_validate_low());
+#endif /* UNIV_DEBUG || UNIV_BUF_DEBUG */
+
+	rbt_free(buf_pool->flush_rbt);
+	buf_pool->flush_rbt = NULL;
+
+	buf_pool_mutex_exit();
+}
+
 /************************************************************************
 Inserts a modified block into the flush list. */
 UNIV_INTERN
@@ -50,6 +186,13 @@ buf_flush_insert_into_flush_list(
 	      || (UT_LIST_GET_FIRST(buf_pool->flush_list)->oldest_modification
 		  <= bpage->oldest_modification));
 
+	/* If we are in the recovery then we need to update the flush
+	red-black tree as well. */
+	if (UNIV_LIKELY_NULL(buf_pool->flush_rbt)) {
+		buf_flush_insert_sorted_into_flush_list(bpage);
+		return;
+	}
+
 	switch (buf_page_get_state(bpage)) {
 	case BUF_BLOCK_ZIP_PAGE:
 		mutex_enter(&buf_pool_zip_mutex);
@@ -120,13 +263,28 @@ buf_flush_insert_sorted_into_flush_list(
 	}
 
 	prev_b = NULL;
+
+	/* For the most part when this function is called the flush_rbt
+	should not be NULL. In a very rare boundary case it is possible
+	that the flush_rbt has already been freed by the recovery thread
+	before the last page was hooked up in the flush_list by the
+	io-handler thread. In that case we'll  just do a simple
+	linear search in the else block. */
+	if (buf_pool->flush_rbt) {
+
+		prev_b = buf_flush_insert_in_flush_rbt(bpage);
+
+	} else {
+
 		b = UT_LIST_GET_FIRST(buf_pool->flush_list);
 
-	while (b && b->oldest_modification > bpage->oldest_modification) {
+		while (b && b->oldest_modification
+		       > bpage->oldest_modification) {
 			ut_ad(b->in_flush_list);
 			prev_b = b;
 			b = UT_LIST_GET_NEXT(list, b);
 		}
+	}
 
 	if (prev_b == NULL) {
 		UT_LIST_ADD_FIRST(list, buf_pool->flush_list, bpage);
@@ -242,6 +400,11 @@ buf_flush_remove(
 		break;
 	}
 
+	/* If the flush_rbt is active then delete from it as well. */
+	if (UNIV_LIKELY_NULL(buf_pool->flush_rbt)) {
+		buf_flush_delete_from_flush_rbt(bpage);
+	}
+
 	bpage->oldest_modification = 0;
 
 	ut_d(UT_LIST_VALIDATE(list, buf_page_t, buf_pool->flush_list));
@@ -1275,6 +1438,15 @@ buf_flush_validate_low(void)
 		ut_a(buf_page_in_file(bpage));
 		ut_a(om > 0);
 
+		/* If we are in recovery mode i.e.: flush_rbt != NULL
+		then each block in the flush_list must also be present
+		in the flush_rbt. */
+		if (UNIV_LIKELY_NULL(buf_pool->flush_rbt)) {
+			ut_a(*rbt_value(buf_page_t*,
+			     rbt_lookup(buf_pool->flush_rbt, &bpage))
+			     == bpage);
+		}
+
 		bpage = UT_LIST_GET_NEXT(list, bpage);
 
 		ut_a(!bpage || om >= bpage->oldest_modification);

buf/buf0rea.c

@@ -745,14 +745,14 @@ buf_read_recv_pages(
 		while (buf_pool->n_pend_reads >= recv_n_pool_free_frames / 2) {
 
 			os_aio_simulated_wake_handler_threads();
-			os_thread_sleep(500000);
+			os_thread_sleep(10000);
 
 			count++;
 
-			if (count > 100) {
+			if (count > 1000) {
 				fprintf(stderr,
 					"InnoDB: Error: InnoDB has waited for"
-					" 50 seconds for pending\n"
+					" 10 seconds for pending\n"
 					"InnoDB: reads to the buffer pool to"
 					" be finished.\n"
 					"InnoDB: Number of pending reads %lu,"

include/buf0buf.h

@@ -30,6 +30,7 @@ Created 11/5/1995 Heikki Tuuri
 #include "sync0rw.h"
 #include "hash0hash.h"
 #include "ut0byte.h"
+#include "ut0rbt.h"
 #include "os0proc.h"
 #include "page0types.h"
 
@@ -1285,6 +1286,19 @@ struct buf_pool_struct{
 					/* this is in the set state when there
 					is no flush batch of the given type
 					running */
+	ib_rbt_t*	flush_rbt;	/* a red-black tree is used
+					exclusively during recovery to
+					speed up insertions in the
+					flush_list. This tree contains
+					blocks in order of
+					oldest_modification LSN and is
+					kept in sync with the
+					flush_list.
+					Each member of the tree MUST
+					also be on the flush_list.
+					This tree is relevant only in
+					recovery and is set to NULL
+					once the recovery is over. */
 	ulint		ulint_clock;	/* a sequence number used to count
 					time. NOTE! This counter wraps
 					around at 4 billion (if ulint ==

include/buf0flu.h

@@ -126,6 +126,22 @@ buf_flush_validate(void);
 		/* out: TRUE if ok */
 #endif /* UNIV_DEBUG || UNIV_BUF_DEBUG */
 
+/************************************************************************
+Initialize the red-black tree to speed up insertions into the flush_list
+during recovery process. Should be called at the start of recovery
+process before any page has been read/written. */
+UNIV_INTERN
+void
+buf_flush_init_flush_rbt(void);
+/*==========================*/
+
+/************************************************************************
+Frees up the red-black tree. */
+UNIV_INTERN
+void
+buf_flush_free_flush_rbt(void);
+/*==========================*/
+
 /* When buf_flush_free_margin is called, it tries to make this many blocks
 available to replacement in the free list and at the end of the LRU list (to
 make sure that a read-ahead batch can be read efficiently in a single

log/log0recv.c

@@ -101,7 +101,9 @@ UNIV_INTERN ulint	recv_max_parsed_page_no		= 0;
 /* This many frames must be left free in the buffer pool when we scan
 the log and store the scanned log records in the buffer pool: we will
 use these free frames to read in pages when we start applying the
-log records to the database. */
+log records to the database.
+This is the default value. If the actual size of the buffer pool is
+larger than 10 MB we'll set this value to 512. */
 
 UNIV_INTERN ulint	recv_n_pool_free_frames		= 256;
 
@@ -156,6 +158,12 @@ recv_sys_init(
 		return;
 	}
 
+	/* Initialize red-black tree for fast insertions into the
+	flush_list during recovery process.
+	As this initialization is done while holding the buffer pool
+	mutex we perform it before acquiring recv_sys->mutex. */
+	buf_flush_init_flush_rbt();
+
 	mutex_enter(&(recv_sys->mutex));
 
 	if (!recover_from_backup) {
@@ -165,6 +173,12 @@ recv_sys_init(
 		recv_is_from_backup = TRUE;
 	}
 
+	/* Set appropriate value of recv_n_pool_free_frames. */
+	if (buf_pool_get_curr_size() >= (10 * 1024 * 1024)) {
+		/* Buffer pool of size greater than 10 MB. */
+		recv_n_pool_free_frames = 512;
+	}
+
 	recv_sys->buf = ut_malloc(RECV_PARSING_BUF_SIZE);
 	recv_sys->len = 0;
 	recv_sys->recovered_offset = 0;
@@ -231,6 +245,9 @@ recv_sys_free(void)
 	recv_sys->heap = NULL;
 
 	mutex_exit(&(recv_sys->mutex));
+
+	/* Free up the flush_rbt. */
+	buf_flush_free_flush_rbt();
 }
 #endif /* UNIV_LOG_DEBUG */