[MySQL 5.6] MySQL 5.6 group commit 性能测试及内部实现流程
尽管Mariadb以及Facebook在long long time ago就fix掉了这个臭名昭著的问题，但官方直到 MySQL5.6 版本才Fix掉，本文主要关注三点：
1.MySQL 5.6的性能如何
2.在5.6中Group commit的三阶段实现流程
新参数
MySQL 5.6提供了两个参数来控制binlog group commit：
单位为微妙，用于从flush队列中取事务的超时时间，这主要是防止并发事务过高，导致某些事务的RT上升。
可以阅读函数MYSQL_BIN_LOG::process_flush_stage_queue 来理解其功能
当设置为0时，事务可能以和binlog不相同的顺序被提交，从下面的测试也可以看出，这会稍微提升点性能，但并不是特别明显.
性能测试
老规矩，先测试看看性能
sysbench, 全内存操作，5个sbtest表，每个表1000000行数据
基本配置：
innodb_flush_log_at_trx_commit=1
table_open_cache_instances=5
metadata_locks_hash_instances = 32
metadata_locks_cache_size=2048
performance_schema_instrument = ‘%=on’
performance_schema=ON
innodb_lru_scan_depth=8192
innodb_purge_threads = 4
关闭Performance Schema consumer：
mysql> update setup_consumers set ENABLED = ‘NO';
Query OK, 4 rows affected (0.02 sec)
Rows matched: 12 Changed: 4 Warnings: 0
sysbench/sysbench –debug=off –test=sysbench/tests/db/update_index.lua –oltp-tables-count=5 –oltp-point-selects=0 –oltp-table-size=1000000 –num-threads=1000 –max-requests=10000000000 –max-time=7200 –oltp-auto-inc=off –mysql-engine-trx=yes –mysql-table-engine=innodb –oltp-test-mod=complex –mysql-db=test –mysql-host=$HOST –mysql-port=3306 –mysql-user=xx run
update_index.lua
| threads | sync_binlog = 0 | sync_binlog = 1 | sync_binlog =1binlog_order_commits=0 |
| 1 | 900 | 610 | 620 |
| 20 | 13,800 | 7,000 | 7,400 |
| 60 | 20,000 | 14,500 | 16,000 |
| 120 | 25,100 | 21,054 | 23,000 |
| 200 | 27,900 | 25,400 | 27,800 |
| 400 | 33,100 | 30,700 | 31,300 |
| 600 | 32,800 | 31,500 | 29,326 |
| 1000 | 20,400 | 20,200 | 20,500 |
我的机器在压到1000个并发时，CPU已经几乎全部耗完。
可以看到，并发度越高，group commit的效果越好，在达到600以上并发时，设置sync_binlog=1或者0已经没有TPS的区别。
但问题是。我们的业务压力很少会达到这么高的压力，低负载下，设置sync_binlog=1依旧增加了单个线程的开销。
另外也观察到，设置binlog_max_flush_queue_time对TPS的影响并不明显。
实现原理
我们知道，binlog和innodb在5.1及以后的版本中采用类似两阶段提交的方式，关于group commit问题的前世今生，可以阅读MATS的博客，讲述的非常详细。嗯，评论也比较有意思。。。。。
以下集中在5.6中binlog如何做group commit。在5.6中，将binlog的commit阶段分为三个阶段：flush stage、sync stage以及commit stage。5.6的实现思路和Mariadb的思路类似，都是维护一个队列，第一个进入该队列的作为leader线程，否则作为follower线程。leader线程收集follower的事务，并负责做sync，follower线程等待leader通知操作完成。
这三个阶段中，每个阶段都会去维护一个队列：
Mutex_queue m_queue[STAGE_COUNTER];
不同session的THD使用the->next_to_commit来链接，实际上，在如下三个阶段，尽管维护了三个队列，但队列中所有的THD实际上都是通过next_to_commit连接起来了。
在binlog的XA_COMMIT阶段(MYSQL_BIN_LOG::commit),完成事务的最后一个xid事件后，，这时候会进入MYSQL_BIN_LOG::ordered_commit，开始3个阶段的流程：
int MYSQL_BIN_LOG::ordered_commit(THD *thd, bool all, bool skip_commit) { DBUG_ENTER("MYSQL_BIN_LOG::ordered_commit"); int flush_error= 0; my_off_t total_bytes= 0; bool do_rotate= false; /* These values are used while flushing a transaction, so clear everything. Notes: - It would be good if we could keep transaction coordinator log-specific data out of the THD structure, but that is not the case right now. - Everything in the transaction structure is reset when calling ha_commit_low since that calls st_transaction::cleanup. */ thd->transaction.flags.pending= true; thd->commit_error= THD::CE_NONE; thd->next_to_commit= NULL; thd->durability_property= HA_IGNORE_DURABILITY; thd->transaction.flags.real_commit= all; thd->transaction.flags.xid_written= false; thd->transaction.flags.commit_low= !skip_commit; thd->transaction.flags.run_hooks= !skip_commit; #ifndef DBUG_OFF /* The group commit Leader may have to wait for follower whose transaction is not ready to be preempted. Initially the status is pessimistic. Preemption guarding logics is necessary only when DBUG_ON is set. It won't be required for the dbug-off case as long as the follower won't execute any thread-specific write access code in this method, which is the case as of current. */ thd->transaction.flags.ready_preempt= 0; #endif DBUG_PRINT("enter", ("flags.pending: %s, commit_error: %d, thread_id: %lu", YESNO(thd->transaction.flags.pending), thd->commit_error, thd->thread_id)); /* Stage #1: flushing transactions to binary log While flushing, we allow new threads to enter and will process them in due time. Once the queue was empty, we cannot reap anything more since it is possible that a thread entered and appointed itself leader for the flush phase. */ if (change_stage(thd, Stage_manager::FLUSH_STAGE, thd, NULL, &LOCK_log)) { DBUG_PRINT("return", ("Thread ID: %lu, commit_error: %d", thd->thread_id, thd->commit_error)); DBUG_RETURN(finish_commit(thd)); } THD *wait_queue= NULL; flush_error= process_flush_stage_queue(&total_bytes, &do_rotate, &wait_queue); my_off_t flush_end_pos= 0; if (flush_error == 0 && total_bytes > 0) flush_error= flush_cache_to_file(&flush_end_pos); /* If the flush finished successfully, we can call the after_flush hook. Being invoked here, we have the guarantee that the hook is executed before the before/after_send_hooks on the dump thread preventing race conditions among these plug-ins. */ if (flush_error == 0) { const char *file_name_ptr= log_file_name + dirname_length(log_file_name); DBUG_ASSERT(flush_end_pos != 0); if (RUN_HOOK(binlog_storage, after_flush, (thd, file_name_ptr, flush_end_pos))) { sql_print_error("Failed to run 'after_flush' hooks"); flush_error= ER_ERROR_ON_WRITE; } signal_update(); DBUG_EXECUTE_IF("crash_commit_after_log", DBUG_SUICIDE();); } /* Stage #2: Syncing binary log file to disk */ bool need_LOCK_log= (get_sync_period() == 1); /* LOCK_log is not released when sync_binlog is 1. It guarantees that the events are not be replicated by dump threads before they are synced to disk. */ if (change_stage(thd, Stage_manager::SYNC_STAGE, wait_queue, need_LOCK_log ? NULL : &LOCK_log, &LOCK_sync)) { DBUG_PRINT("return", ("Thread ID: %lu, commit_error: %d", thd->thread_id, thd->commit_error)); DBUG_RETURN(finish_commit(thd)); } THD *final_queue= stage_manager.fetch_queue_for(Stage_manager::SYNC_STAGE); if (flush_error == 0 && total_bytes > 0) { DEBUG_SYNC(thd, "before_sync_binlog_file"); std::pair<bool, bool> result= sync_binlog_file(false); flush_error= result.first; } if (need_LOCK_log) mysql_mutex_unlock(&LOCK_log); /* Stage #3: Commit all transactions in order. This stage is skipped if we do not need to order the commits and each thread have to execute the handlerton commit instead. Howver, since we are keeping the lock from the previous stage, we need to unlock it if we skip the stage. */ if (opt_binlog_order_commits) { if (change_stage(thd, Stage_manager::COMMIT_STAGE, final_queue, &LOCK_sync, &LOCK_commit)) { DBUG_PRINT("return", ("Thread ID: %lu, commit_error: %d", thd->thread_id, thd->commit_error)); DBUG_RETURN(finish_commit(thd)); } THD *commit_queue= stage_manager.fetch_queue_for(Stage_manager::COMMIT_STAGE); DBUG_EXECUTE_IF("semi_sync_3-way_deadlock", DEBUG_SYNC(thd, "before_process_commit_stage_queue");); process_commit_stage_queue(thd, commit_queue); mysql_mutex_unlock(&LOCK_commit); /* Process after_commit after LOCK_commit is released for avoiding 3-way deadlock among user thread, rotate thread and dump thread. */ process_after_commit_stage_queue(thd, commit_queue); final_queue= commit_queue; } else mysql_mutex_unlock(&LOCK_sync); /* Commit done so signal all waiting threads */ stage_manager.signal_done(final_queue); /* Finish the commit before executing a rotate, or run the risk of a deadlock. We don't need the return value here since it is in thd->commit_error, which is returned below. */ (void) finish_commit(thd); /* If we need to rotate, we do it without commit error. Otherwise the thd->commit_error will be possibly reset. */ if (do_rotate && thd->commit_error == THD::CE_NONE) { /* Do not force the rotate as several consecutive groups may request unnecessary rotations. NOTE: Run purge_logs wo/ holding LOCK_log because it does not need the mutex. Otherwise causes various deadlocks. */ DEBUG_SYNC(thd, "ready_to_do_rotation"); bool check_purge= false; mysql_mutex_lock(&LOCK_log); int error= rotate(false, &check_purge); mysql_mutex_unlock(&LOCK_log); if (error) thd->commit_error= THD::CE_COMMIT_ERROR; else if (check_purge) purge(); } DBUG_RETURN(thd->commit_error); }
###flush stage
int MYSQL_BIN_LOG::process_flush_stage_queue(my_off_t *total_bytes_var, bool *rotate_var, THD **out_queue_var) { DBUG_ASSERT(total_bytes_var && rotate_var && out_queue_var); my_off_t total_bytes= 0; int flush_error= 1; mysql_mutex_assert_owner(&LOCK_log); my_atomic_rwlock_rdlock(&opt_binlog_max_flush_queue_time_lock); const ulonglong max_udelay= my_atomic_load32(&opt_binlog_max_flush_queue_time); my_atomic_rwlock_rdunlock(&opt_binlog_max_flush_queue_time_lock); const ulonglong start_utime= max_udelay > 0 ? my_micro_time() : 0; /* First we read the queue until it either is empty or the difference between the time we started and the current time is too large. We remember the first thread we unqueued, because this will be the beginning of the out queue. */ bool has_more= true; THD *first_seen= NULL; while ((max_udelay == 0 || my_micro_time() < start_utime + max_udelay) && has_more) { std::pair<bool,THD*> current= stage_manager.pop_front(Stage_manager::FLUSH_STAGE); std::pair<int,my_off_t> result= flush_thread_caches(current.second); has_more= current.first; total_bytes+= result.second; if (flush_error == 1) flush_error= result.first; if (first_seen == NULL) first_seen= current.second; } /* Either the queue is empty, or we ran out of time. If we ran out of time, we have to fetch the entire queue (and flush it) since otherwise the next batch will not have a leader. */ if (has_more) { THD *queue= stage_manager.fetch_queue_for(Stage_manager::FLUSH_STAGE); for (THD *head= queue ; head ; head = head->next_to_commit) { std::pair<int,my_off_t> result= flush_thread_caches(head); total_bytes+= result.second; if (flush_error == 1) flush_error= result.first; } if (first_seen == NULL) first_seen= queue; } *out_queue_var= first_seen; *total_bytes_var= total_bytes; if (total_bytes > 0 && my_b_tell(&log_file) >= (my_off_t) max_size) *rotate_var= true; return flush_error; }
change_stage(thd, Stage_manager::FLUSH_STAGE, thd, NULL, &LOCK_log)
|–>stage_manager.enroll_for(stage, queue, leave_mutex) //将当前线程加入到m_queue[FLUSH_STAGE]中，如果是队列的第一个线程，就被设置为leader，否则就是follower线程，线程会这其中睡眠，直到被leader唤醒(m_cond_done)
|–>leader线程持有LOCK_log锁，从change_state线程返回false.
flush_error= process_flush_stage_queue(&total_bytes, &do_rotate, &wait_queue); //只有leader线程才会进入这个逻辑
|–>首先读取队列，直到队列为空，或者超时（超时时间是通过参数binlog_max_flush_queue_time来控制)为止，对读到的每个线程做flush_thread_caches，将binlog刷到cache中。注意在出队列的时候，可能还有新的session被append到队列中，设置超时的目的也正在于此
|–>如果是超时，这时候队列中还有session的话，就取出整个队列的头部线程，并将原队列置空(fetch_queue_for)，然后对取出的session进行flush_thread_caches
|–>判断总的写入binlog的byte数是否超过max bin log size，如果超过了，就设置rotate标记
flush_error= flush_cache_to_file(&flush_end_pos);
|–>将I/O Cache中的内容写到文件中
signal_update() //通知dump线程有新的Binlog
###sync stage
change_stage(thd, Stage_manager::SYNC_STAGE, wait_queue, &LOCK_log, &LOCK_sync)
|–>stage_manager.enroll_for(stage, queue, leave_mutex) //当前线程加入到m_queue[SYNC_STAGE]队列中，释放lock_log锁；同样的如果是SYNC_STAGE队列的leader，则立刻返回，否则进行condition wait.
|–>leader线程加上Lock_sync锁
final_queue= stage_manager.fetch_queue_for(Stage_manager::SYNC_STAGE); //从SYNC_STAGE队列中取出来，并清空队列，主要用于commit阶段
std::pair<bool, bool> result= sync_binlog_file(false); //刷binlog 文件(如果设置了sync_binlog的话)
简单的理解就是，在flush stage阶段形成N批的组session，在SYNC阶段又会由这N批组产生出新的leader来负责做最耗时的sync操作
###commit stage
commit阶段受到参数binlog_order_commits限制
当binlog_order_commits关闭时，直接unlock LOCK_sync，由各个session自行进入Innodb commit阶段(随后调用的finish_commit(thd))，这样不会保证binlog和事务commit的顺序一致，如果你不关注innodb的ibdata中记录的binlog信息，那么可以关闭这个选项来稍微提高点性能
当打开binlog_order_commits时，才会进入commit stage，如下描述的
change_stage(thd, Stage_manager::COMMIT_STAGE,final_queue, &LOCK_sync, &LOCK_commit)
|–>进入新的COMMIT_STAGE队列，释放LOCK_sync锁，新的leader获取LOCK_commit锁，其他的session等待
THD *commit_queue= stage_manager.fetch_queue_for(Stage_manager::COMMIT_STAGE); //取出并清空COMMIT_STAGE队列
process_commit_stage_queue(thd, commit_queue, flush_error)
|–>这里会遍历所有的线程，然后调用ha_commit_low->innobase_commit进入innodb层依次提交
完成上述步骤后，解除LOCK_commit锁
stage_manager.signal_done(final_queue);
|–>将所有Pending的线程的标记置为false（thd->transaction.flags.pending= false）并做m_cond_done广播，唤醒pending的线程
(void) finish_commit(the); //如果binlog_order_commits设置为FALSE，就会进入这一步来提交存储引擎层事务; 另外还会更新grid信息
Innodb的group commit和mariadb的类似，都只有两次sync，即在prepare阶段sync，以及sync Binlog文件（双一配置），为了保证rotate时，所有前一个binlog的事件的redo log都被刷到磁盘，会在函数new_file_impl中调用如下代码段：
if (DBUG_EVALUATE_IF(“expire_logs_always”, 0, 1)
&& (error= ha_flush_logs(NULL)))
goto end;
ha_flush_logs 会调用存储引擎接口刷日志文件
参考文档
http://dimitrik.free.fr/blog/archives/2012/06/mysql-performance-binlog-group-commit-in-56.html
http://mysqlmusings.blogspot.com/2012/06/binary-log-group-commit-in-mysql-56.html
MySQL 5.6.10 source code