MYSQL-45: InnoDB should allow flushing to be gated by age of pages

Currently using innodb_max_dirty_pages_pct especially with a large
innodb_io_capacity does not enforce any lower limit on the number of
dirty pages, and this can increase the number of writes dramatically.
While normally this would be OK (flush if you can) this is not
desirable behavior on SSDs where write lifetime is limited.

In order to reduce and control the frequency of write requests to
SSD, this change introduces a new variable named
innodb_flush_dirty_pages_age which can be used to set the minimum age
of dirty pages to be flushed from the buffer pool if the number of
dirty pages is below the maximum percentage. This impacts the
frequency with which frequently modified pages are flushed, which can
help reduce IO load in systems with frequently modified pages (hot
spots).

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

@@ -0,0 +1,56 @@
+SET @start_global_value = @@global.innodb_flush_dirty_pages_age;
+SELECT @start_global_value;
+@start_global_value
+0
+Valid values are zero or above
+select @@global.innodb_flush_dirty_pages_age >=0;
+@@global.innodb_flush_dirty_pages_age >=0
+1
+select @@global.innodb_flush_dirty_pages_age;
+@@global.innodb_flush_dirty_pages_age
+0
+select @@session.innodb_flush_dirty_pages_age;
+ERROR HY000: Variable 'innodb_flush_dirty_pages_age' is a GLOBAL variable
+show global variables like 'innodb_flush_dirty_pages_age';
+Variable_name	Value
+innodb_flush_dirty_pages_age	0
+show session variables like 'innodb_flush_dirty_pages_age';
+Variable_name	Value
+innodb_flush_dirty_pages_age	0
+select * from information_schema.global_variables where variable_name='innodb_flush_dirty_pages_age';
+VARIABLE_NAME	VARIABLE_VALUE
+INNODB_FLUSH_DIRTY_PAGES_AGE	0
+select * from information_schema.session_variables where variable_name='innodb_flush_dirty_pages_age';
+VARIABLE_NAME	VARIABLE_VALUE
+INNODB_FLUSH_DIRTY_PAGES_AGE	0
+set global innodb_flush_dirty_pages_age=10;
+select @@global.innodb_flush_dirty_pages_age;
+@@global.innodb_flush_dirty_pages_age
+10
+select * from information_schema.global_variables where variable_name='innodb_flush_dirty_pages_age';
+VARIABLE_NAME	VARIABLE_VALUE
+INNODB_FLUSH_DIRTY_PAGES_AGE	10
+select * from information_schema.session_variables where variable_name='innodb_flush_dirty_pages_age';
+VARIABLE_NAME	VARIABLE_VALUE
+INNODB_FLUSH_DIRTY_PAGES_AGE	10
+set session innodb_flush_dirty_pages_age=1;
+ERROR HY000: Variable 'innodb_flush_dirty_pages_age' is a GLOBAL variable and should be set with SET GLOBAL
+set global innodb_flush_dirty_pages_age=1.1;
+ERROR 42000: Incorrect argument type to variable 'innodb_flush_dirty_pages_age'
+set global innodb_flush_dirty_pages_age=1e1;
+ERROR 42000: Incorrect argument type to variable 'innodb_flush_dirty_pages_age'
+set global innodb_flush_dirty_pages_age="foo";
+ERROR 42000: Incorrect argument type to variable 'innodb_flush_dirty_pages_age'
+set global innodb_flush_dirty_pages_age=-7;
+Warnings:
+Warning	1292	Truncated incorrect innodb_flush_dirty_pages_age value: '-7'
+select @@global.innodb_flush_dirty_pages_age;
+@@global.innodb_flush_dirty_pages_age
+0
+select * from information_schema.global_variables where variable_name='innodb_flush_dirty_pages_age';
+VARIABLE_NAME	VARIABLE_VALUE
+INNODB_FLUSH_DIRTY_PAGES_AGE	0
+SET @@global.innodb_flush_dirty_pages_age = @start_global_value;
+SELECT @@global.innodb_flush_dirty_pages_age;
+@@global.innodb_flush_dirty_pages_age
+0

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

@@ -0,0 +1,52 @@
+
+#
+# 2010-01-25 - Added
+#
+
+--source include/have_innodb.inc
+
+SET @start_global_value = @@global.innodb_flush_dirty_pages_age;
+SELECT @start_global_value;
+
+#
+# exists as global only
+#
+--echo Valid values are zero or above
+select @@global.innodb_flush_dirty_pages_age >=0;
+select @@global.innodb_flush_dirty_pages_age;
+--error ER_INCORRECT_GLOBAL_LOCAL_VAR
+select @@session.innodb_flush_dirty_pages_age;
+show global variables like 'innodb_flush_dirty_pages_age';
+show session variables like 'innodb_flush_dirty_pages_age';
+select * from information_schema.global_variables where variable_name='innodb_flush_dirty_pages_age';
+select * from information_schema.session_variables where variable_name='innodb_flush_dirty_pages_age';
+
+#
+# show that it's writable
+#
+set global innodb_flush_dirty_pages_age=10;
+select @@global.innodb_flush_dirty_pages_age;
+select * from information_schema.global_variables where variable_name='innodb_flush_dirty_pages_age';
+select * from information_schema.session_variables where variable_name='innodb_flush_dirty_pages_age';
+--error ER_GLOBAL_VARIABLE
+set session innodb_flush_dirty_pages_age=1;
+
+#
+# incorrect types
+#
+--error ER_WRONG_TYPE_FOR_VAR
+set global innodb_flush_dirty_pages_age=1.1;
+--error ER_WRONG_TYPE_FOR_VAR
+set global innodb_flush_dirty_pages_age=1e1;
+--error ER_WRONG_TYPE_FOR_VAR
+set global innodb_flush_dirty_pages_age="foo";
+
+set global innodb_flush_dirty_pages_age=-7;
+select @@global.innodb_flush_dirty_pages_age;
+select * from information_schema.global_variables where variable_name='innodb_flush_dirty_pages_age';
+
+#
+# cleanup
+#
+SET @@global.innodb_flush_dirty_pages_age = @start_global_value;
+SELECT @@global.innodb_flush_dirty_pages_age;

storage/innobase/handler/ha_innodb.cc

@@ -11683,6 +11683,13 @@ static MYSQL_SYSVAR_UINT(trx_rseg_n_slots_debug, trx_rseg_n_slots_debug,
   NULL, NULL, 0, 0, 1024, 0);
 #endif /* UNIV_DEBUG */
 
+static MYSQL_SYSVAR_UINT(flush_dirty_pages_age, srv_buf_flush_dirty_pages_age,
+  PLUGIN_VAR_RQCMDARG,
+  "Minimum age (since the last flush) of pages to be flushed from "
+  "the buffer pool if the number of dirty pages is below the maximum "
+  "percentage.",
+  NULL, NULL, 0, 0, UINT_MAX32, 0);
+
 static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(additional_mem_pool_size),
   MYSQL_SYSVAR(autoextend_increment),
@@ -11756,6 +11763,7 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
 #ifdef UNIV_DEBUG
   MYSQL_SYSVAR(trx_rseg_n_slots_debug),
 #endif /* UNIV_DEBUG */
+  MYSQL_SYSVAR(flush_dirty_pages_age),
   NULL
 };
 

storage/innobase/include/srv0srv.h

@@ -162,6 +162,8 @@ extern ulint	srv_buf_pool_curr_size;	/*!< current size in bytes */
 extern ulint	srv_mem_pool_size;
 extern ulint	srv_lock_table_size;
 
+extern uint	srv_buf_flush_dirty_pages_age;
+
 extern ulint	srv_n_file_io_threads;
 extern my_bool	srv_random_read_ahead;
 extern ulong	srv_read_ahead_threshold;

storage/innobase/srv/srv0srv.c

@@ -218,6 +218,8 @@ UNIV_INTERN ulint	srv_buf_pool_curr_size	= 0;
 UNIV_INTERN ulint	srv_mem_pool_size	= ULINT_MAX;
 UNIV_INTERN ulint	srv_lock_table_size	= ULINT_MAX;
 
+UNIV_INTERN uint	srv_buf_flush_dirty_pages_age = 0;
+
 /* This parameter is deprecated. Use srv_n_io_[read|write]_threads
 instead. */
 UNIV_INTERN ulint	srv_n_file_io_threads	= ULINT_MAX;
@@ -509,6 +511,9 @@ thread ensures that we flush the log files at least once per
 second. */
 static time_t	srv_last_log_flush_time;
 
+/* Records the time of the last flush limit calculation. */
+static time_t	srv_last_flush_limit_time;
+
 /* The master thread performs various tasks based on the current
 state of IO activity and the level of IO utilization is past
 intervals. Following macros define thresholds for these conditions. */
@@ -2668,6 +2673,8 @@ srv_master_thread(
 {
 	buf_pool_stat_t buf_stat;
 	srv_slot_t*	slot;
+	ib_uint64_t	flush_lsn_limit = 0;
+	ib_uint64_t	new_flush_lsn_limit = srv_start_lsn;
 	ulint		old_activity_count;
 	ulint		n_pages_purged	= 0;
 	ulint		n_bytes_merged;
@@ -2844,7 +2851,9 @@ srv_master_thread(
 #endif
 	/* If i/os during the 10 second period were less than 200% of
 	capacity, we assume that there is free disk i/o capacity
-	available, and it makes sense to flush srv_io_capacity pages.
+	available, and it makes sense to flush srv_io_capacity pages
+	unless there is a need to reduce the frequency of write requests
+	to the storage device (which might be the case for SSDs).
 
 	Note that this is done regardless of the fraction of dirty
 	pages relative to the max requested by the user. The one second
@@ -2856,12 +2865,33 @@ srv_master_thread(
 	n_ios = log_sys->n_log_ios + buf_stat.n_pages_read
 		+ buf_stat.n_pages_written;
 
+	/* Constantly record the youngest age at which a dirty page
+	might be flushed. */
+	if (difftime(time(NULL), srv_last_flush_limit_time)
+	    > srv_buf_flush_dirty_pages_age) {
+
+		/* The limit prevents the flushing of pages dirtied within
+		the last srv_buf_flush_dirty_pages_age seconds. */
+		flush_lsn_limit = new_flush_lsn_limit;
+		new_flush_lsn_limit = log_get_lsn();
+		srv_last_flush_limit_time = time(NULL);
+	}
+
+	/* If not limited by age, try to keep the number of modified pages
+	in the buffer pool as low as possible. */
+	if (!srv_buf_flush_dirty_pages_age) {
+		flush_lsn_limit = IB_ULONGLONG_MAX;
+	}
+
 	srv_main_10_second_loops++;
 	if (n_pend_ios < SRV_PEND_IO_THRESHOLD
-	    && (n_ios - n_ios_very_old < SRV_PAST_IO_ACTIVITY)) {
+	    && (n_ios - n_ios_very_old < SRV_PAST_IO_ACTIVITY)
+	    && flush_lsn_limit) {
 
 		srv_main_thread_op_info = "flushing buffer pool pages";
-		buf_flush_list(PCT_IO(100), IB_ULONGLONG_MAX);
+		buf_flush_list(PCT_IO(100), flush_lsn_limit);
+
+		flush_lsn_limit = 0;
 
 		/* Flush logs if needed */
 		srv_sync_log_buffer_in_background();