Bug 12356373 - PERFORMANCE REGRESSION FROM 5.1 TO 5.5 : GROUP BY:

The title of the bug is a little confusing. The actual fix is to
reintroduce random readahead inside InnoDB with a dynamic, global
switch innodb_random_read_ahead [default = off].

Approved by: Sunny Bains
rb://696

storage/innodb_plugin/ChangeLog

+2011-07-19	The InnoDB Team
+
+	* buf/buf0buf.c, buf/buf0rea.c, handler/ha_innodb.cc,
+	include/buf0buf.h, include/buf0buf.ic, include/srv0srv.h,
+	srv/srv0srv.c:
+	Fix bug#Bug 12356373 by reintroducing random readahead
+
 2011-06-30	The InnoDB Team
 
 	* row/row0row.c:

storage/innodb_plugin/buf/buf0buf.c

@@ -3590,12 +3590,16 @@ buf_print_io(
 
 	/* Statistics about read ahead algorithm */
 	fprintf(file, "Pages read ahead %.2f/s,"
-		" evicted without access %.2f/s\n",
+		" evicted without access %.2f/s,"
+		" Random read ahead %.2f/s\n",
 		(buf_pool->stat.n_ra_pages_read
 		- buf_pool->old_stat.n_ra_pages_read)
 		/ time_elapsed,
 		(buf_pool->stat.n_ra_pages_evicted
 		- buf_pool->old_stat.n_ra_pages_evicted)
+		/ time_elapsed,
+		(buf_pool->stat.n_ra_pages_read_rnd
+		- buf_pool->old_stat.n_ra_pages_read_rnd)
 		/ time_elapsed);
 
 	/* Print some values to help us with visualizing what is

storage/innodb_plugin/buf/buf0rea.c

@@ -38,6 +38,14 @@ Created 11/5/1995 Heikki Tuuri
 #include "srv0start.h"
 #include "srv0srv.h"
 
+/** The size in blocks of the area where the random read-ahead algorithm counts
+the accessed pages when deciding whether to read-ahead */
+#define	BUF_READ_AHEAD_RANDOM_AREA	BUF_READ_AHEAD_AREA
+
+/** There must be at least this many pages in buf_pool in the area to start
+a random read-ahead */
+#define BUF_READ_AHEAD_RANDOM_THRESHOLD	(5 + BUF_READ_AHEAD_RANDOM_AREA / 8)
+
 /** The linear read-ahead area size */
 #define	BUF_READ_AHEAD_LINEAR_AREA	BUF_READ_AHEAD_AREA
 
@@ -157,6 +165,165 @@ buf_read_page_low(
 	return(1);
 }
 
+/********************************************************************//**
+Applies a random read-ahead in buf_pool if there are at least a threshold
+value of accessed pages from the random read-ahead area. Does not read any
+page, not even the one at the position (space, offset), if the read-ahead
+mechanism is not activated. NOTE 1: the calling thread may own latches on
+pages: to avoid deadlocks this function must be written such that it cannot
+end up waiting for these latches! NOTE 2: the calling thread must want
+access to the page given: this rule is set to prevent unintended read-aheads
+performed by ibuf routines, a situation which could result in a deadlock if
+the OS does not support asynchronous i/o.
+@return number of page read requests issued; NOTE that if we read ibuf
+pages, it may happen that the page at the given page number does not
+get read even if we return a positive value! */
+static
+ulint
+buf_read_ahead_random(
+/*==================*/
+	ulint	space,	/*!< in: space id */
+	ulint	zip_size,/*!< in: compressed page size in bytes, or 0 */
+	ulint	offset)	/*!< in: page number of a page which the current thread
+			wants to access */
+{
+	ib_int64_t	tablespace_version;
+	ulint		recent_blocks	= 0;
+	ulint		count;
+	ulint		ibuf_mode;
+	ulint		low, high;
+	ulint		err;
+	ulint		i;
+	ulint		buf_read_ahead_random_area;
+
+	if (!srv_random_read_ahead) {
+		/* Disabled by user */
+		return(0);
+	}
+
+	if (srv_startup_is_before_trx_rollback_phase) {
+		/* No read-ahead to avoid thread deadlocks */
+		return(0);
+	}
+
+	if (ibuf_bitmap_page(zip_size, offset)
+	    || trx_sys_hdr_page(space, offset)) {
+
+		/* If it is an ibuf bitmap page or trx sys hdr, we do
+		no read-ahead, as that could break the ibuf page access
+		order */
+
+		return(0);
+	}
+
+	/* Remember the tablespace version before we ask the tablespace size
+	below: if DISCARD + IMPORT changes the actual .ibd file meanwhile, we
+	do not try to read outside the bounds of the tablespace! */
+
+	tablespace_version = fil_space_get_version(space);
+
+	buf_read_ahead_random_area = BUF_READ_AHEAD_RANDOM_AREA;
+
+	low  = (offset / buf_read_ahead_random_area)
+		* buf_read_ahead_random_area;
+	high = (offset / buf_read_ahead_random_area + 1)
+		* buf_read_ahead_random_area;
+	if (high > fil_space_get_size(space)) {
+
+		high = fil_space_get_size(space);
+	}
+
+	buf_pool_mutex_enter();
+
+	if (buf_pool->n_pend_reads
+	    > buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) {
+		buf_pool_mutex_exit();
+
+		return(0);
+	}
+
+	/* Count how many blocks in the area have been recently accessed,
+	that is, reside near the start of the LRU list. */
+
+	for (i = low; i < high; i++) {
+		const buf_page_t*	bpage = buf_page_hash_get(space, i);
+
+		if (bpage
+		    && buf_page_is_accessed(bpage)
+		    && buf_page_peek_if_young(bpage)) {
+
+			recent_blocks++;
+
+			if (recent_blocks >= BUF_READ_AHEAD_RANDOM_THRESHOLD) {
+
+				buf_pool_mutex_exit();
+				goto read_ahead;
+			}
+		}
+	}
+
+	buf_pool_mutex_exit();
+	/* Do nothing */
+	return(0);
+
+read_ahead:
+	/* Read all the suitable blocks within the area */
+
+	if (ibuf_inside()) {
+		ibuf_mode = BUF_READ_IBUF_PAGES_ONLY;
+	} else {
+		ibuf_mode = BUF_READ_ANY_PAGE;
+	}
+
+	count = 0;
+
+	for (i = low; i < high; i++) {
+		/* It is only sensible to do read-ahead in the non-sync aio
+		mode: hence FALSE as the first parameter */
+
+		if (!ibuf_bitmap_page(zip_size, i)) {
+			count += buf_read_page_low(
+				&err, FALSE,
+				ibuf_mode | OS_AIO_SIMULATED_WAKE_LATER,
+				space, zip_size, FALSE,
+				tablespace_version, i);
+			if (err == DB_TABLESPACE_DELETED) {
+				ut_print_timestamp(stderr);
+				fprintf(stderr,
+					"  InnoDB: Warning: in random"
+					" readahead trying to access\n"
+					"InnoDB: tablespace %lu page %lu,\n"
+					"InnoDB: but the tablespace does not"
+					" exist or is just being dropped.\n",
+					(ulong) space, (ulong) i);
+			}
+		}
+	}
+
+	/* In simulated aio we wake the aio handler threads only after
+	queuing all aio requests, in native aio the following call does
+	nothing: */
+
+	os_aio_simulated_wake_handler_threads();
+
+#ifdef UNIV_DEBUG
+	if (buf_debug_prints && (count > 0)) {
+		fprintf(stderr,
+			"Random read-ahead space %lu offset %lu pages %lu\n",
+			(ulong) space, (ulong) offset,
+			(ulong) count);
+	}
+#endif /* UNIV_DEBUG */
+
+	/* Read ahead is considered one I/O operation for the purpose of
+	LRU policy decision. */
+	buf_LRU_stat_inc_io();
+
+	buf_pool->stat.n_ra_pages_read_rnd += count;
+	return(count);
+}
+
+
 /********************************************************************//**
 High-level function which reads a page asynchronously from a file to the
 buffer buf_pool if it is not already there. Sets the io_fix flag and sets
@@ -175,6 +342,9 @@ buf_read_page(
 	ulint		count;
 	ulint		err;
 
+	count = buf_read_ahead_random(space, zip_size, offset);
+	srv_buf_pool_reads += count;
+
 	tablespace_version = fil_space_get_version(space);
 
 	/* We do the i/o in the synchronous aio mode to save thread

storage/innodb_plugin/handler/ha_innodb.cc

@@ -501,6 +501,8 @@ static SHOW_VAR innodb_status_variables[]= {
   (char*) &export_vars.innodb_buffer_pool_pages_misc,	  SHOW_LONG},
   {"buffer_pool_pages_total",
   (char*) &export_vars.innodb_buffer_pool_pages_total,	  SHOW_LONG},
+  {"buffer_pool_read_ahead_rnd",
+  (char*) &export_vars.innodb_buffer_pool_read_ahead_rnd, SHOW_LONG},
   {"buffer_pool_read_ahead",
   (char*) &export_vars.innodb_buffer_pool_read_ahead,	  SHOW_LONG},
   {"buffer_pool_read_ahead_evicted",
@@ -11027,6 +11029,11 @@ static MYSQL_SYSVAR_UINT(change_buffering_debug, ibuf_debug,
   NULL, NULL, 0, 0, 1, 0);
 #endif /* UNIV_DEBUG || UNIV_IBUF_DEBUG */
 
+static MYSQL_SYSVAR_BOOL(random_read_ahead, srv_random_read_ahead,
+  PLUGIN_VAR_NOCMDARG,
+  "Whether to use read ahead for random access within an extent.",
+  NULL, NULL, FALSE);
+
 static MYSQL_SYSVAR_ULONG(read_ahead_threshold, srv_read_ahead_threshold,
   PLUGIN_VAR_RQCMDARG,
   "Number of pages that must be accessed sequentially for InnoDB to "
@@ -11091,6 +11098,7 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
 #if defined UNIV_DEBUG || defined UNIV_IBUF_DEBUG
   MYSQL_SYSVAR(change_buffering_debug),
 #endif /* UNIV_DEBUG || UNIV_IBUF_DEBUG */
+  MYSQL_SYSVAR(random_read_ahead),
   MYSQL_SYSVAR(read_ahead_threshold),
   MYSQL_SYSVAR(io_capacity),
   NULL

storage/innodb_plugin/include/buf0buf.h

@@ -427,6 +427,18 @@ buf_block_get_freed_page_clock(
 	__attribute__((pure));
 
 /********************************************************************//**
+Tells if a block is still close enough to the MRU end of the LRU list
+meaning that it is not in danger of getting evicted and also implying
+that it has been accessed recently.
+Note that this is for heuristics only and does not reserve buffer pool
+mutex.
+@return	TRUE if block is close to MRU end of LRU */
+UNIV_INLINE
+ibool
+buf_page_peek_if_young(
+/*===================*/
+	const buf_page_t*	bpage);	/*!< in: block */
+/********************************************************************//**
 Recommends a move of a block to the start of the LRU list if there is danger
 of dropping from the buffer pool. NOTE: does not reserve the buffer pool
 mutex.
@@ -1334,6 +1346,8 @@ struct buf_pool_stat_struct{
 	ulint	n_pages_written;/*!< number write operations */
 	ulint	n_pages_created;/*!< number of pages created
 				in the pool with no read */
+	ulint	n_ra_pages_read_rnd;/*!< number of pages read in
+				as part of random read ahead */
 	ulint	n_ra_pages_read;/*!< number of pages read in
 				as part of read ahead */
 	ulint	n_ra_pages_evicted;/*!< number of read ahead

storage/innodb_plugin/include/buf0buf.ic

@@ -61,6 +61,27 @@ buf_block_get_freed_page_clock(
 	return(buf_page_get_freed_page_clock(&block->page));
 }
 
+/********************************************************************//**
+Tells if a block is still close enough to the MRU end of the LRU list
+meaning that it is not in danger of getting evicted and also implying
+that it has been accessed recently.
+Note that this is for heuristics only and does not reserve buffer pool
+mutex.
+@return	TRUE if block is close to MRU end of LRU */
+UNIV_INLINE
+ibool
+buf_page_peek_if_young(
+/*===================*/
+	const buf_page_t*	bpage)	/*!< in: block */
+{
+	/* FIXME: bpage->freed_page_clock is 31 bits */
+	return((buf_pool->freed_page_clock & ((1UL << 31) - 1))
+	       < ((ulint) bpage->freed_page_clock
+		  + (buf_pool->curr_size
+		     * (BUF_LRU_OLD_RATIO_DIV - buf_LRU_old_ratio)
+		     / (BUF_LRU_OLD_RATIO_DIV * 4))));
+}
+
 /********************************************************************//**
 Recommends a move of a block to the start of the LRU list if there is danger
 of dropping from the buffer pool. NOTE: does not reserve the buffer pool
@@ -89,12 +110,7 @@ buf_page_peek_if_too_old(
 		buf_pool->stat.n_pages_not_made_young++;
 		return(FALSE);
 	} else {
-		/* FIXME: bpage->freed_page_clock is 31 bits */
-		return((buf_pool->freed_page_clock & ((1UL << 31) - 1))
-		       > ((ulint) bpage->freed_page_clock
-			  + (buf_pool->curr_size
-			     * (BUF_LRU_OLD_RATIO_DIV - buf_LRU_old_ratio)
-			     / (BUF_LRU_OLD_RATIO_DIV * 4))));
+		return(!buf_page_peek_if_young(bpage));
 	}
 }
 

storage/innodb_plugin/include/srv0srv.h

@@ -143,6 +143,7 @@ extern ulint	srv_mem_pool_size;
 extern ulint	srv_lock_table_size;
 
 extern ulint	srv_n_file_io_threads;
+extern my_bool	srv_random_read_ahead;
 extern ulong	srv_read_ahead_threshold;
 extern ulint	srv_n_read_io_threads;
 extern ulint	srv_n_write_io_threads;
@@ -618,6 +619,7 @@ struct export_var_struct{
 	ulint innodb_buffer_pool_wait_free;	/*!< srv_buf_pool_wait_free */
 	ulint innodb_buffer_pool_pages_flushed;	/*!< srv_buf_pool_flushed */
 	ulint innodb_buffer_pool_write_requests;/*!< srv_buf_pool_write_requests */
+	ulint innodb_buffer_pool_read_ahead_rnd;/*!< srv_read_ahead_rnd */
 	ulint innodb_buffer_pool_read_ahead;	/*!< srv_read_ahead */
 	ulint innodb_buffer_pool_read_ahead_evicted;/*!< srv_read_ahead evicted*/
 	ulint innodb_dblwr_pages_written;	/*!< srv_dblwr_pages_written */

storage/innodb_plugin/srv/srv0srv.c

@@ -203,6 +203,8 @@ UNIV_INTERN ulint	srv_n_file_io_threads	= ULINT_MAX;
 UNIV_INTERN ulint	srv_n_read_io_threads	= ULINT_MAX;
 UNIV_INTERN ulint	srv_n_write_io_threads	= ULINT_MAX;
 
+/* Switch to enable random read ahead. */
+UNIV_INTERN my_bool	srv_random_read_ahead	= FALSE;
 /* User settable value of the number of pages that must be present
 in the buffer cache and accessed sequentially for InnoDB to trigger a
 readahead request. */
@@ -1906,6 +1908,8 @@ srv_export_innodb_status(void)
 	export_vars.innodb_buffer_pool_wait_free = srv_buf_pool_wait_free;
 	export_vars.innodb_buffer_pool_pages_flushed = srv_buf_pool_flushed;
 	export_vars.innodb_buffer_pool_reads = srv_buf_pool_reads;
+	export_vars.innodb_buffer_pool_read_ahead_rnd
+		= buf_pool->stat.n_ra_pages_read_rnd;
 	export_vars.innodb_buffer_pool_read_ahead
 		= buf_pool->stat.n_ra_pages_read;
 	export_vars.innodb_buffer_pool_read_ahead_evicted