Merge pull request #125 from grooverdan/MDEV-8923_innodb_buffer_pool_dump_pct

MDEV-8923: port innodb_buffer_pool_dump_pct from MySQL

WL#6504 InnoDB buffer pool dump/load enchantments

This patch consists of two parts:

1. Dump only the hottest N% of the buffer pool(s)
2. Prevent hogging the server duing BP load

From MySQL - commit b409342c43ce2edb68807100a77001367c7e6b8e

Add testcases for innodb_buffer_pool_dump_pct.

Part of the code authored by Daniel Black.

mysql-test/suite/sys_vars/r/innodb_buffer_pool_dump_pct.result

+SET @orig = @@global.innodb_buffer_pool_dump_pct;
+SELECT @orig;
+@orig
+100
+SET GLOBAL innodb_buffer_pool_dump_pct=3, GLOBAL innodb_buffer_pool_dump_now = ON;
+SET GLOBAL innodb_buffer_pool_dump_pct=0;
+SELECT @@global.innodb_buffer_pool_dump_pct;
+@@global.innodb_buffer_pool_dump_pct
+1
+SHOW WARNINGS;
+Level	Code	Message
+Warning	1292	Truncated incorrect innodb_buffer_pool_dump_pct value: '0'
+SET GLOBAL innodb_buffer_pool_dump_pct=101;
+SELECT @@global.innodb_buffer_pool_dump_pct;
+@@global.innodb_buffer_pool_dump_pct
+100
+SHOW WARNINGS;
+Level	Code	Message
+Warning	1292	Truncated incorrect innodb_buffer_pool_dump_pct value: '101'
+SET GLOBAL innodb_buffer_pool_dump_pct=@orig;

mysql-test/suite/sys_vars/t/innodb_buffer_pool_dump_pct.test

+#
+# Basic test for innodb_buffer_pool_dump_pct
+#
+
+-- source include/have_innodb.inc
+
+# Check the default value
+SET @orig = @@global.innodb_buffer_pool_dump_pct;
+SELECT @orig;
+
+# Do the dump
+SET GLOBAL innodb_buffer_pool_dump_pct=3, GLOBAL innodb_buffer_pool_dump_now = ON;
+
+# Wait for the dump to complete
+let $wait_condition =
+  SELECT SUBSTR(variable_value, 1, 33) = 'Buffer pool(s) dump completed at '
+  FROM information_schema.global_status
+  WHERE LOWER(variable_name) = 'innodb_buffer_pool_dump_status';
+-- source include/wait_condition.inc
+
+# Confirm that the dump file has been created
+-- let $file = `SELECT CONCAT(@@datadir, @@global.innodb_buffer_pool_filename)`
+-- file_exists $file
+
+--disable_warnings
+SET GLOBAL innodb_buffer_pool_dump_pct=0;
+SELECT @@global.innodb_buffer_pool_dump_pct;
+SHOW WARNINGS;
+SET GLOBAL innodb_buffer_pool_dump_pct=101;
+SELECT @@global.innodb_buffer_pool_dump_pct;
+SHOW WARNINGS;
+--enable_warnings
+
+SET GLOBAL innodb_buffer_pool_dump_pct=@orig;

storage/innobase/buf/buf0dump.cc

@@ -228,6 +228,16 @@ buf_dump(
 			continue;
 		}
 
+		if (srv_buf_pool_dump_pct != 100) {
+			ut_ad(srv_buf_pool_dump_pct < 100);
+
+			n_pages = n_pages * srv_buf_pool_dump_pct / 100;
+
+			if (n_pages == 0) {
+				n_pages = 1;
+			}
+		}
+
 		dump = static_cast<buf_dump_t*>(
 			ut_malloc(n_pages * sizeof(*dump))) ;
 
@@ -242,9 +252,9 @@ buf_dump(
 			return;
 		}
 
-		for (bpage = UT_LIST_GET_LAST(buf_pool->LRU), j = 0;
-		     bpage != NULL;
-		     bpage = UT_LIST_GET_PREV(LRU, bpage), j++) {
+		for (bpage = UT_LIST_GET_FIRST(buf_pool->LRU), j = 0;
+		     bpage != NULL && j < n_pages;
+		     bpage = UT_LIST_GET_NEXT(LRU, bpage), j++) {
 
 			ut_a(buf_page_in_file(bpage));
 
@@ -368,6 +378,72 @@ buf_dump_sort(
 			      buf_dump_cmp);
 }
 
+/*****************************************************************//**
+Artificially delay the buffer pool loading if necessary. The idea of
+this function is to prevent hogging the server with IO and slowing down
+too much normal client queries. */
+UNIV_INLINE
+void
+buf_load_throttle_if_needed(
+/*========================*/
+	ulint*	last_check_time,	/*!< in/out: miliseconds since epoch
+					of the last time we did check if
+					throttling is needed, we do the check
+					every srv_io_capacity IO ops. */
+	ulint*	last_activity_count,
+	ulint	n_io)			/*!< in: number of IO ops done since
+					buffer pool load has started */
+{
+	if (n_io % srv_io_capacity < srv_io_capacity - 1) {
+		return;
+	}
+
+	if (*last_check_time == 0 || *last_activity_count == 0) {
+		*last_check_time = ut_time_ms();
+		*last_activity_count = srv_get_activity_count();
+		return;
+	}
+
+	/* srv_io_capacity IO operations have been performed by buffer pool
+	load since the last time we were here. */
+
+	/* If no other activity, then keep going without any delay. */
+	if (srv_get_activity_count() == *last_activity_count) {
+		return;
+	}
+
+	/* There has been other activity, throttle. */
+
+	ulint	now = ut_time_ms();
+	ulint	elapsed_time = now - *last_check_time;
+
+	/* Notice that elapsed_time is not the time for the last
+	srv_io_capacity IO operations performed by BP load. It is the
+	time elapsed since the last time we detected that there has been
+	other activity. This has a small and acceptable deficiency, e.g.:
+	1. BP load runs and there is no other activity.
+	2. Other activity occurs, we run N IO operations after that and
+	   enter here (where 0 <= N < srv_io_capacity).
+	3. last_check_time is very old and we do not sleep at this time, but
+	   only update last_check_time and last_activity_count.
+	4. We run srv_io_capacity more IO operations and call this function
+	   again.
+	5. There has been more other activity and thus we enter here.
+	6. Now last_check_time is recent and we sleep if necessary to prevent
+	   more than srv_io_capacity IO operations per second.
+	The deficiency is that we could have slept at 3., but for this we
+	would have to update last_check_time before the
+	"cur_activity_count == *last_activity_count" check and calling
+	ut_time_ms() that often may turn out to be too expensive. */
+
+	if (elapsed_time < 1000 /* 1 sec (1000 mili secs) */) {
+		os_thread_sleep((1000 - elapsed_time) * 1000 /* micro secs */);
+	}
+
+	*last_check_time = ut_time_ms();
+	*last_activity_count = srv_get_activity_count();
+}
+
 /*****************************************************************//**
 Perform a buffer pool load from the file specified by
 innodb_buffer_pool_filename. If any errors occur then the value of
@@ -529,6 +605,9 @@ buf_load()
 
 	ut_free(dump_tmp);
 
+	ulint	last_check_time = 0;
+	ulint	last_activity_cnt = 0;
+
 	for (i = 0; i < dump_n && !SHUTTING_DOWN(); i++) {
 
 		buf_read_page_async(BUF_DUMP_SPACE(dump[i]),
@@ -552,6 +631,9 @@ buf_load()
 				"Buffer pool(s) load aborted on request");
 			return;
 		}
+
+		buf_load_throttle_if_needed(
+			&last_check_time, &last_activity_cnt, i);
 	}
 
 	ut_free(dump);

storage/innobase/handler/ha_innodb.cc

@@ -18787,6 +18787,11 @@ static MYSQL_SYSVAR_BOOL(buffer_pool_dump_at_shutdown, srv_buffer_pool_dump_at_s
   "Dump the buffer pool into a file named @@innodb_buffer_pool_filename",
   NULL, NULL, FALSE);
 
+static MYSQL_SYSVAR_ULONG(buffer_pool_dump_pct, srv_buf_pool_dump_pct,
+  PLUGIN_VAR_RQCMDARG,
+  "Dump only the hottest N% of each buffer pool, defaults to 100",
+  NULL, NULL, 100, 1, 100, 0);
+
 #ifdef UNIV_DEBUG
 static MYSQL_SYSVAR_STR(buffer_pool_evict, srv_buffer_pool_evict,
   PLUGIN_VAR_RQCMDARG,
@@ -19524,6 +19529,7 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(buffer_pool_filename),
   MYSQL_SYSVAR(buffer_pool_dump_now),
   MYSQL_SYSVAR(buffer_pool_dump_at_shutdown),
+  MYSQL_SYSVAR(buffer_pool_dump_pct),
 #ifdef UNIV_DEBUG
   MYSQL_SYSVAR(buffer_pool_evict),
 #endif /* UNIV_DEBUG */

storage/innobase/include/srv0srv.h

@@ -380,6 +380,8 @@ extern ulong	srv_flush_neighbors;	/*!< whether or not to flush
 					neighbors of a block */
 extern ulint	srv_buf_pool_old_size;	/*!< previously requested size */
 extern ulint	srv_buf_pool_curr_size;	/*!< current size in bytes */
+extern ulong	srv_buf_pool_dump_pct;	/*!< dump that may % of each buffer
+					pool during BP dump */
 extern ulint	srv_mem_pool_size;
 extern ulint	srv_lock_table_size;
 

storage/innobase/srv/srv0srv.cc

@@ -263,6 +263,8 @@ UNIV_INTERN ulong	srv_flush_neighbors	= 1;
 UNIV_INTERN ulint	srv_buf_pool_old_size;
 /* current size in kilobytes */
 UNIV_INTERN ulint	srv_buf_pool_curr_size	= 0;
+/* dump that may % of each buffer pool during BP dump */
+UNIV_INTERN ulong	srv_buf_pool_dump_pct;
 /* size in bytes */
 UNIV_INTERN ulint	srv_mem_pool_size	= ULINT_MAX;
 UNIV_INTERN ulint	srv_lock_table_size	= ULINT_MAX;

storage/xtradb/buf/buf0dump.cc

@@ -228,6 +228,16 @@ buf_dump(
 			continue;
 		}
 
+		if (srv_buf_pool_dump_pct != 100) {
+			ut_ad(srv_buf_pool_dump_pct < 100);
+
+			n_pages = n_pages * srv_buf_pool_dump_pct / 100;
+
+			if (n_pages == 0) {
+				n_pages = 1;
+			}
+		}
+
 		dump = static_cast<buf_dump_t*>(
 			ut_malloc(n_pages * sizeof(*dump))) ;
 
@@ -242,9 +252,9 @@ buf_dump(
 			return;
 		}
 
-		for (bpage = UT_LIST_GET_LAST(buf_pool->LRU), j = 0;
-		     bpage != NULL;
-		     bpage = UT_LIST_GET_PREV(LRU, bpage), j++) {
+		for (bpage = UT_LIST_GET_FIRST(buf_pool->LRU), j = 0;
+		     bpage != NULL && j < n_pages;
+		     bpage = UT_LIST_GET_NEXT(LRU, bpage), j++) {
 
 			ut_a(buf_page_in_file(bpage));
 
@@ -368,6 +378,72 @@ buf_dump_sort(
 			      buf_dump_cmp);
 }
 
+/*****************************************************************//**
+Artificially delay the buffer pool loading if necessary. The idea of
+this function is to prevent hogging the server with IO and slowing down
+too much normal client queries. */
+UNIV_INLINE
+void
+buf_load_throttle_if_needed(
+/*========================*/
+	ulint*	last_check_time,	/*!< in/out: miliseconds since epoch
+					of the last time we did check if
+					throttling is needed, we do the check
+					every srv_io_capacity IO ops. */
+	ulint*	last_activity_count,
+	ulint	n_io)			/*!< in: number of IO ops done since
+					buffer pool load has started */
+{
+	if (n_io % srv_io_capacity < srv_io_capacity - 1) {
+		return;
+	}
+
+	if (*last_check_time == 0 || *last_activity_count == 0) {
+		*last_check_time = ut_time_ms();
+		*last_activity_count = srv_get_activity_count();
+		return;
+	}
+
+	/* srv_io_capacity IO operations have been performed by buffer pool
+	load since the last time we were here. */
+
+	/* If no other activity, then keep going without any delay. */
+	if (srv_get_activity_count() == *last_activity_count) {
+		return;
+	}
+
+	/* There has been other activity, throttle. */
+
+	ulint	now = ut_time_ms();
+	ulint	elapsed_time = now - *last_check_time;
+
+	/* Notice that elapsed_time is not the time for the last
+	srv_io_capacity IO operations performed by BP load. It is the
+	time elapsed since the last time we detected that there has been
+	other activity. This has a small and acceptable deficiency, e.g.:
+	1. BP load runs and there is no other activity.
+	2. Other activity occurs, we run N IO operations after that and
+	   enter here (where 0 <= N < srv_io_capacity).
+	3. last_check_time is very old and we do not sleep at this time, but
+	   only update last_check_time and last_activity_count.
+	4. We run srv_io_capacity more IO operations and call this function
+	   again.
+	5. There has been more other activity and thus we enter here.
+	6. Now last_check_time is recent and we sleep if necessary to prevent
+	   more than srv_io_capacity IO operations per second.
+	The deficiency is that we could have slept at 3., but for this we
+	would have to update last_check_time before the
+	"cur_activity_count == *last_activity_count" check and calling
+	ut_time_ms() that often may turn out to be too expensive. */
+
+	if (elapsed_time < 1000 /* 1 sec (1000 mili secs) */) {
+		os_thread_sleep((1000 - elapsed_time) * 1000 /* micro secs */);
+	}
+
+	*last_check_time = ut_time_ms();
+	*last_activity_count = srv_get_activity_count();
+}
+
 /*****************************************************************//**
 Perform a buffer pool load from the file specified by
 innodb_buffer_pool_filename. If any errors occur then the value of
@@ -529,6 +605,9 @@ buf_load()
 
 	ut_free(dump_tmp);
 
+	ulint	last_check_time = 0;
+	ulint	last_activity_cnt = 0;
+
 	for (i = 0; i < dump_n && !SHUTTING_DOWN(); i++) {
 
 		buf_read_page_async(BUF_DUMP_SPACE(dump[i]),
@@ -552,6 +631,9 @@ buf_load()
 				"Buffer pool(s) load aborted on request");
 			return;
 		}
+
+		buf_load_throttle_if_needed(
+			&last_check_time, &last_activity_cnt, i);
 	}
 
 	ut_free(dump);

storage/xtradb/handler/ha_innodb.cc

@@ -19934,6 +19934,11 @@ static MYSQL_SYSVAR_BOOL(buffer_pool_dump_at_shutdown, srv_buffer_pool_dump_at_s
   "Dump the buffer pool into a file named @@innodb_buffer_pool_filename",
   NULL, NULL, FALSE);
 
+static MYSQL_SYSVAR_ULONG(buffer_pool_dump_pct, srv_buf_pool_dump_pct,
+  PLUGIN_VAR_RQCMDARG,
+  "Dump only the hottest N% of each buffer pool, defaults to 100",
+  NULL, NULL, 100, 1, 100, 0);
+
 #ifdef UNIV_DEBUG
 static MYSQL_SYSVAR_STR(buffer_pool_evict, srv_buffer_pool_evict,
   PLUGIN_VAR_RQCMDARG,
@@ -20736,6 +20741,7 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(buffer_pool_filename),
   MYSQL_SYSVAR(buffer_pool_dump_now),
   MYSQL_SYSVAR(buffer_pool_dump_at_shutdown),
+  MYSQL_SYSVAR(buffer_pool_dump_pct),
 #ifdef UNIV_DEBUG
   MYSQL_SYSVAR(buffer_pool_evict),
 #endif /* UNIV_DEBUG */

storage/xtradb/include/srv0srv.h

@@ -403,6 +403,8 @@ extern ulong	srv_flush_neighbors;	/*!< whether or not to flush
 					neighbors of a block */
 extern ulint	srv_buf_pool_old_size;	/*!< previously requested size */
 extern ulint	srv_buf_pool_curr_size;	/*!< current size in bytes */
+extern ulong	srv_buf_pool_dump_pct;	/*!< dump that may % of each buffer
+					pool during BP dump */
 extern ulint	srv_mem_pool_size;
 extern ulint	srv_lock_table_size;
 

storage/xtradb/srv/srv0srv.cc

@@ -296,6 +296,8 @@ UNIV_INTERN ulong	srv_flush_neighbors	= 1;
 UNIV_INTERN ulint	srv_buf_pool_old_size;
 /* current size in kilobytes */
 UNIV_INTERN ulint	srv_buf_pool_curr_size	= 0;
+/* dump that may % of each buffer pool during BP dump */
+UNIV_INTERN ulong	srv_buf_pool_dump_pct;
 /* size in bytes */
 UNIV_INTERN ulint	srv_mem_pool_size	= ULINT_MAX;
 UNIV_INTERN ulint	srv_lock_table_size	= ULINT_MAX;