Merge pull request #248 from sensssz/10.2-vats

MDEV-11039 - Add new scheduling algorithm for reducing tail latencies (for 10.2)

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

@@ -1447,6 +1447,20 @@ NUMERIC_BLOCK_SIZE	NULL
 ENUM_VALUE_LIST	OFF,ON
 READ_ONLY	YES
 COMMAND_LINE_ARGUMENT	NONE
+VARIABLE_NAME	INNODB_LOCK_SCHEDULE_ALGORITHM
+SESSION_VALUE	NULL
+GLOBAL_VALUE	fcfs
+GLOBAL_VALUE_ORIGIN	COMPILE-TIME
+DEFAULT_VALUE	fcfs
+VARIABLE_SCOPE	GLOBAL
+VARIABLE_TYPE	ENUM
+VARIABLE_COMMENT	The algorithm Innodb uses for deciding which locks to grant next when a lock is released. Possible values are FCFS grant the locks in First-Come-First-Served order; VATS use the Variance-Aware-Transaction-Scheduling algorithm, which uses an Eldest-Transaction-First heuristic.
+NUMERIC_MIN_VALUE	NULL
+NUMERIC_MAX_VALUE	NULL
+NUMERIC_BLOCK_SIZE	NULL
+ENUM_VALUE_LIST	fcfs,vats
+READ_ONLY	NO
+COMMAND_LINE_ARGUMENT	REQUIRED
 VARIABLE_NAME	INNODB_LOCK_WAIT_TIMEOUT
 SESSION_VALUE	50
 GLOBAL_VALUE	50

storage/innobase/handler/ha_innodb.cc

@@ -379,6 +379,22 @@ static TYPELIB innodb_default_row_format_typelib = {
 	NULL
 };
 
+/** Possible values of the parameter innodb_lock_schedule_algorithm */
+static const char* innodb_lock_schedule_algorithm_names[] = {
+	"fcfs",
+	"vats",
+	NullS
+};
+
+/** Used to define an enumerate type of the system variable
+innodb_lock_schedule_algorithm. */
+static TYPELIB innodb_lock_schedule_algorithm_typelib = {
+	array_elements(innodb_lock_schedule_algorithm_names) - 1,
+	"innodb_lock_schedule_algorithm_typelib",
+	innodb_lock_schedule_algorithm_names,
+	NULL
+};
+
 /* The following counter is used to convey information to InnoDB
 about server activity: in case of normal DML ops it is not
 sensible to call srv_active_wake_master_thread after each
@@ -1671,7 +1687,7 @@ thd_is_replication_slave_thread(
 /*============================*/
 	THD*	thd)	/*!< in: thread handle */
 {
-	return((ibool) thd_slave_thread(thd));
+	return thd && ((ibool) thd_slave_thread(thd));
 }
 
 /******************************************************************//**
@@ -22783,6 +22799,18 @@ static MYSQL_SYSVAR_ULONG(doublewrite_batch_size, srv_doublewrite_batch_size,
   NULL, NULL, 120, 1, 127, 0);
 #endif /* defined UNIV_DEBUG || defined UNIV_PERF_DEBUG */
 
+static MYSQL_SYSVAR_ENUM(lock_schedule_algorithm, innodb_lock_schedule_algorithm,
+  PLUGIN_VAR_RQCMDARG,
+  "The algorithm Innodb uses for deciding which locks to grant next when"
+  " a lock is released. Possible values are"
+  " FCFS"
+  " grant the locks in First-Come-First-Served order;"
+  " VATS"
+  " use the Variance-Aware-Transaction-Scheduling algorithm, which"
+  " uses an Eldest-Transaction-First heuristic.",
+  NULL, NULL, INNODB_LOCK_SCHEDULE_ALGORITHM_FCFS,
+  &innodb_lock_schedule_algorithm_typelib);
+
 static MYSQL_SYSVAR_ULONG(buffer_pool_instances, srv_buf_pool_instances,
   PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
   "Number of buffer pool instances, set to higher value on high-end machines to increase scalability",
@@ -23656,6 +23684,7 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(ft_sort_pll_degree),
   MYSQL_SYSVAR(large_prefix),
   MYSQL_SYSVAR(force_load_corrupted),
+  MYSQL_SYSVAR(lock_schedule_algorithm),
   MYSQL_SYSVAR(locks_unsafe_for_binlog),
   MYSQL_SYSVAR(lock_wait_timeout),
   MYSQL_SYSVAR(page_size),

storage/innobase/include/lock0lock.h

@@ -40,6 +40,17 @@ Created 5/7/1996 Heikki Tuuri
 #include "gis0rtree.h"
 #include "lock0prdt.h"
 
+/** Alternatives for innodb_lock_schedule_algorithm, which can be changed by
+	setting innodb_lock_schedule_algorithm. */
+enum innodb_lock_schedule_algorithm_t {
+	/*!< First Come First Served */
+	INNODB_LOCK_SCHEDULE_ALGORITHM_FCFS,
+	/*!< Variance-Aware-Transaction-Scheduling */
+	INNODB_LOCK_SCHEDULE_ALGORITHM_VATS
+};
+
+extern ulong innodb_lock_schedule_algorithm;
+
 // Forward declaration
 class ReadView;
 

storage/innobase/include/trx0trx.h

@@ -1096,6 +1096,8 @@ struct trx_t {
 
 	time_t		start_time;	/*!< time the state last time became
 					TRX_STATE_ACTIVE */
+	clock_t		start_time_micro; /*!< start time of the transaction
+					in microseconds. */
 	lsn_t		commit_lsn;	/*!< lsn at the time of the commit */
 	table_id_t	table_id;	/*!< Table to drop iff dict_operation
 					== TRX_DICT_OP_TABLE, or 0. */

storage/innobase/trx/trx0trx.cc

@@ -1484,6 +1484,7 @@ trx_start_low(
 	} else {
 		trx->start_time = ut_time();
 	}
+	trx->start_time_micro = clock();
 
 	ut_a(trx->error_state == DB_SUCCESS);