背景：
MySQL 5.5開始新增一個數據庫：PERFORMANCE_SCHEMA，主要用於收集數據庫服務器性能參數。並且庫里表的存儲引擎均為PERFORMANCE_SCHEMA，而用戶是不能創建存儲引擎為PERFORMANCE_SCHEMA的表。MySQL5.5默認是關閉的，需要手動開啟，在配置文件里添加：
[mysqld] performance_schema=ON
查看是否開啟：
mysql>show variables like 'performance_schema'; +--------------------+-------+ | Variable_name | Value | +--------------------+-------+ | performance_schema | ON | +--------------------+-------+
從MySQL5.6開始，默認打開，本文就從MySQL5.6來說明，在數據庫使用當中PERFORMANCE_SCHEMA的一些比較常用的功能。具體的信息可以查看官方文檔。
相關表信息：
一：配置（setup）表：
zjy@performance_schema 10:16:56>show tables like '%setup%'; +----------------------------------------+ | Tables_in_performance_schema (%setup%) | +----------------------------------------+ | setup_actors | | setup_consumers | | setup_instruments | | setup_objects | | setup_timers | +----------------------------------------+
1，setup_actors：配置用戶緯度的監控，默認監控所有用戶。
zjy@performance_schema 10:19:11>select * from setup_actors; +------+------+------+ | HOST | USER | ROLE | +------+------+------+ | % | % | % | +------+------+------+
2，setup_consumers：配置events的消費者類型，即收集的events寫入到哪些統計表中。
zjy@: performance_schema 10:23:35>select * from setup_consumers; +--------------------------------+---------+ | NAME | ENABLED | +--------------------------------+---------+ | events_stages_current | NO | | events_stages_history | NO | | events_stages_history_long | NO | | events_statements_current | YES | | events_statements_history | NO | | events_statements_history_long | NO | | events_waits_current | NO | | events_waits_history | NO | | events_waits_history_long | NO | | global_instrumentation | YES | | thread_instrumentation | YES | | statements_digest | YES | +--------------------------------+---------+
這里需要說明的是需要查看哪個就更新其ENABLED列為YES。如：
zjy@performance_schema 10:25:02>update setup_consumers set ENABLED='YES' where NAME in ('events_stages_current','events_waits_current'); Query OK, 2 rows affected (0.00 sec)
更新完后立即生效，但是服務器重啟之后又會變回默認值，要永久生效需要在配置文件里添加：
[mysqld] #performance_schema performance_schema_consumer_events_waits_current=on performance_schema_consumer_events_stages_current=on performance_schema_consumer_events_statements_current=on performance_schema_consumer_events_waits_history=on performance_schema_consumer_events_stages_history=on performance_schema_consumer_events_statements_history=on
即在這些表的前面加上：performance_schema_consumer_xxx。表setup_consumers里面的值有個層級關系：
global_instrumentation > thread_instrumentation = statements_digest > events_stages_current = events_statements_current = events_waits_current > events_stages_history = events_statements_history = events_waits_history > events_stages_history_long = events_statements_history_long = events_waits_history_long
只有上一層次的為YES，才會繼續檢查該本層為YES or NO。global_instrumentation是最高級別consumer，如果它設置為NO，則所有的consumer都會忽略。其中history和history_long存的是current表的歷史記錄條數，history表記錄了每個線程最近等待的10個事件，而history_long表則記錄了最近所有線程產生的10000個事件，這里的10和10000都是可以配置的。這三個表表結構相同，history和history_long表數據都來源於current表。長度通過控制參數：
zjy@performance_schema 11:10:03>show variables like 'performance_schema%history%size'; +--------------------------------------------------------+-------+ | Variable_name | Value | +--------------------------------------------------------+-------+ | performance_schema_events_stages_history_long_size | 10000 | | performance_schema_events_stages_history_size | 10 | | performance_schema_events_statements_history_long_size | 10000 | | performance_schema_events_statements_history_size | 10 | | performance_schema_events_waits_history_long_size | 10000 | | performance_schema_events_waits_history_size | 10 | +--------------------------------------------------------+-------+
3，setup_instruments：配置具體的instrument，主要包含4大類：idle、stage/xxx、statement/xxx、wait/xxx：
zjy@performance_schema 10:56:35>select name,count(*) from setup_instruments group by LEFT(name,5); +---------------------------------+----------+ | name | count(*) | +---------------------------------+----------+ | idle | 1 | | stage/sql/After create | 111 | | statement/sql/select | 179 | | wait/synch/mutex/sql/PAGE::lock | 296 | +---------------------------------+----------+
idle表示socket空閑的時間，stage類表示語句的每個執行階段的統計，statement類統計語句維度的信息，wait類統計各種等待事件，比如IO，mutux，spin_lock,condition等。
4，setup_objects：配置監控對象，默認對mysql，performance_schema和information_schema中的表都不監控，而其它DB的所有表都監控。
zjy@performance_schema 11:00:18>select * from setup_objects; +-------------+--------------------+-------------+---------+-------+ | OBJECT_TYPE | OBJECT_SCHEMA | OBJECT_NAME | ENABLED | TIMED | +-------------+--------------------+-------------+---------+-------+ | TABLE | mysql | % | NO | NO | | TABLE | performance_schema | % | NO | NO | | TABLE | information_schema | % | NO | NO | | TABLE | % | % | YES | YES | +-------------+--------------------+-------------+---------+-------+
5，setup_timers：配置每種類型指令的統計時間單位。MICROSECOND表示統計單位是微妙，CYCLE表示統計單位是時鍾周期，時間度量與CPU的主頻有關，NANOSECOND表示統計單位是納秒。但無論采用哪種度量單位，最終統計表中統計的時間都會裝換到皮秒。（1秒＝1000000000000皮秒）
zjy@performance_schema 11:05:12>select * from setup_timers; +-----------+-------------+ | NAME | TIMER_NAME | +-----------+-------------+ | idle | MICROSECOND | | wait | CYCLE | | stage | NANOSECOND | | statement | NANOSECOND | +-----------+-------------+
二：instance表
1，cond_instances：條件等待對象實例
表中記錄了系統中使用的條件變量的對象，OBJECT_INSTANCE_BEGIN為對象的內存地址。
2，file_instances：文件實例
表中記錄了系統中打開了文件的對象，包括ibdata文件，redo文件，binlog文件，用戶的表文件等，open_count顯示當前文件打開的數目，如果重來沒有打開過，不會出現在表中。
zjy@performance_schema 11:20:04>select * from file_instances limit 2,5; +---------------------------------+--------------------------------------+------------+ | FILE_NAME | EVENT_NAME | OPEN_COUNT | +---------------------------------+--------------------------------------+------------+ | /var/lib/mysql/mysql/plugin.frm | wait/io/file/sql/FRM | 0 | | /var/lib/mysql/mysql/plugin.MYI | wait/io/file/myisam/kfile | 1 | | /var/lib/mysql/mysql/plugin.MYD | wait/io/file/myisam/dfile | 1 | | /var/lib/mysql/ibdata1 | wait/io/file/innodb/innodb_data_file | 2 | | /var/lib/mysql/ib_logfile0 | wait/io/file/innodb/innodb_log_file | 2 | +---------------------------------+--------------------------------------+------------+
3，mutex_instances：互斥同步對象實例
表中記錄了系統中使用互斥量對象的所有記錄，其中name為：wait/synch/mutex/*。LOCKED_BY_THREAD_ID顯示哪個線程正持有mutex，若沒有線程持有，則為NULL。
4，rwlock_instances： 讀寫鎖同步對象實例
表中記錄了系統中使用讀寫鎖對象的所有記錄，其中name為 wait/synch/rwlock/*。WRITE_LOCKED_BY_THREAD_ID為正在持有該對象的thread_id，若沒有線程持有，則為NULL。READ_LOCKED_BY_COUNT為記錄了同時有多少個讀者持有讀鎖。（通過 events_waits_current 表可以知道，哪個線程在等待鎖；通過rwlock_instances知道哪個線程持有鎖。rwlock_instances的缺陷是，只能記錄持有寫鎖的線程，對於讀鎖則無能為力）。
5，socket_instances：活躍會話對象實例
表中記錄了thread_id,socket_id,ip和port，其它表可以通過thread_id與socket_instance進行關聯，獲取IP-PORT信息，能夠與應用對接起來。
event_name主要包含3類：
wait/io/socket/sql/server_unix_socket，服務端unix監聽socket
wait/io/socket/sql/server_tcpip_socket，服務端tcp監聽socket
wait/io/socket/sql/client_connection，客戶端socket
三：Wait表
1，events_waits_current：記錄了當前線程等待的事件
2，events_waits_history：記錄了每個線程最近等待的10個事件
3，events_waits_history_long：記錄了最近所有線程產生的10000個事件
表結構定義如下：
CREATE TABLE `events_waits_current` ( `THREAD_ID` bigint(20) unsigned NOT NULL COMMENT '線程ID', `EVENT_ID` bigint(20) unsigned NOT NULL COMMENT '當前線程的事件ID，和THREAD_ID確定唯一', `END_EVENT_ID` bigint(20) unsigned DEFAULT NULL COMMENT '當事件開始時，這一列被設置為NULL。當事件結束時，再更新為當前的事件ID', `EVENT_NAME` varchar(128) NOT NULL COMMENT '事件名稱', `SOURCE` varchar(64) DEFAULT NULL COMMENT '該事件產生時的源碼文件', `TIMER_START` bigint(20) unsigned DEFAULT NULL COMMENT '事件開始時間（皮秒）', `TIMER_END` bigint(20) unsigned DEFAULT NULL COMMENT '事件結束結束時間（皮秒）', `TIMER_WAIT` bigint(20) unsigned DEFAULT NULL COMMENT '事件等待時間（皮秒）', `SPINS` int(10) unsigned DEFAULT NULL COMMENT '', `OBJECT_SCHEMA` varchar(64) DEFAULT NULL COMMENT '庫名', `OBJECT_NAME` varchar(512) DEFAULT NULL COMMENT '文件名、表名、IP:SOCK值', `OBJECT_TYPE` varchar(64) DEFAULT NULL COMMENT 'FILE、TABLE、TEMPORARY TABLE', `INDEX_NAME` varchar(64) DEFAULT NULL COMMENT '索引名', `OBJECT_INSTANCE_BEGIN` bigint(20) unsigned NOT NULL COMMENT '內存地址', `NESTING_EVENT_ID` bigint(20) unsigned DEFAULT NULL COMMENT '該事件對應的父事件ID', `NESTING_EVENT_TYPE` enum('STATEMENT','STAGE','WAIT') DEFAULT NULL COMMENT '父事件類型(STATEMENT, STAGE, WAIT)', `OPERATION` varchar(32) NOT NULL COMMENT '操作類型（lock, read, write）', `NUMBER_OF_BYTES` bigint(20) DEFAULT NULL COMMENT '', `FLAGS` int(10) unsigned DEFAULT NULL COMMENT '標記' ) ENGINE=PERFORMANCE_SCHEMA DEFAULT CHARSET=utf8
四：Stage 表
1，events_stages_current：記錄了當前線程所處的執行階段
2，events_stages_history：記錄了當前線程所處的執行階段10條歷史記錄
3，events_stages_history_long：記錄了當前線程所處的執行階段10000條歷史記錄
表結構定義如下：
CREATE TABLE `events_stages_current` ( `THREAD_ID` bigint(20) unsigned NOT NULL COMMENT '線程ID', `EVENT_ID` bigint(20) unsigned NOT NULL COMMENT '事件ID', `END_EVENT_ID` bigint(20) unsigned DEFAULT NULL COMMENT '結束事件ID', `EVENT_NAME` varchar(128) NOT NULL COMMENT '事件名稱', `SOURCE` varchar(64) DEFAULT NULL COMMENT '源碼位置', `TIMER_START` bigint(20) unsigned DEFAULT NULL COMMENT '事件開始時間（皮秒）', `TIMER_END` bigint(20) unsigned DEFAULT NULL COMMENT '事件結束結束時間（皮秒）', `TIMER_WAIT` bigint(20) unsigned DEFAULT NULL COMMENT '事件等待時間（皮秒）', `NESTING_EVENT_ID` bigint(20) unsigned DEFAULT NULL COMMENT '該事件對應的父事件ID', `NESTING_EVENT_TYPE` enum('STATEMENT','STAGE','WAIT') DEFAULT NULL COMMENT '父事件類型(STATEMENT, STAGE, WAIT)' ) ENGINE=PERFORMANCE_SCHEMA DEFAULT CHARSET=utf8
五：Statement 表
1，events_statements_current：通過 thread_id+event_id可以唯一確定一條記錄。Statments表只記錄最頂層的請求，SQL語句或是COMMAND，每條語句一行。event_name形式為statement/sql/*，或statement/com/*
2，events_statements_history
3，events_statements_history_long
表結構定義如下：
CREATE TABLE `events_statements_current` ( `THREAD_ID` bigint(20) unsigned NOT NULL COMMENT '線程ID', `EVENT_ID` bigint(20) unsigned NOT NULL COMMENT '事件ID', `END_EVENT_ID` bigint(20) unsigned DEFAULT NULL COMMENT '結束事件ID', `EVENT_NAME` varchar(128) NOT NULL COMMENT '事件名稱', `SOURCE` varchar(64) DEFAULT NULL COMMENT '源碼位置', `TIMER_START` bigint(20) unsigned DEFAULT NULL COMMENT '事件開始時間（皮秒）', `TIMER_END` bigint(20) unsigned DEFAULT NULL COMMENT '事件結束結束時間（皮秒）', `TIMER_WAIT` bigint(20) unsigned DEFAULT NULL COMMENT '事件等待時間（皮秒）', `LOCK_TIME` bigint(20) unsigned NOT NULL COMMENT '鎖時間', `SQL_TEXT` longtext COMMENT '記錄SQL語句', `DIGEST` varchar(32) DEFAULT NULL COMMENT '對SQL_TEXT做MD5產生的32位字符串', `DIGEST_TEXT` longtext COMMENT '將語句中值部分用問號代替，用於SQL語句歸類', `CURRENT_SCHEMA` varchar(64) DEFAULT NULL COMMENT '默認的數據庫名', `OBJECT_TYPE` varchar(64) DEFAULT NULL COMMENT '保留字段', `OBJECT_SCHEMA` varchar(64) DEFAULT NULL COMMENT '保留字段', `OBJECT_NAME` varchar(64) DEFAULT NULL COMMENT '保留字段', `OBJECT_INSTANCE_BEGIN` bigint(20) unsigned DEFAULT NULL COMMENT '內存地址', `MYSQL_ERRNO` int(11) DEFAULT NULL COMMENT '', `RETURNED_SQLSTATE` varchar(5) DEFAULT NULL COMMENT '', `MESSAGE_TEXT` varchar(128) DEFAULT NULL COMMENT '信息', `ERRORS` bigint(20) unsigned NOT NULL COMMENT '錯誤數目', `WARNINGS` bigint(20) unsigned NOT NULL COMMENT '警告數目', `ROWS_AFFECTED` bigint(20) unsigned NOT NULL COMMENT '影響的數目', `ROWS_SENT` bigint(20) unsigned NOT NULL COMMENT '返回的記錄數', `ROWS_EXAMINED` bigint(20) unsigned NOT NULL COMMENT '讀取掃描的記錄數目', `CREATED_TMP_DISK_TABLES` bigint(20) unsigned NOT NULL COMMENT '創建磁盤臨時表數目', `CREATED_TMP_TABLES` bigint(20) unsigned NOT NULL COMMENT '創建臨時表數目', `SELECT_FULL_JOIN` bigint(20) unsigned NOT NULL COMMENT 'join時，第一個表為全表掃描的數目', `SELECT_FULL_RANGE_JOIN` bigint(20) unsigned NOT NULL COMMENT '引用表采用range方式掃描的數目', `SELECT_RANGE` bigint(20) unsigned NOT NULL COMMENT 'join時，第一個表采用range方式掃描的數目', `SELECT_RANGE_CHECK` bigint(20) unsigned NOT NULL COMMENT '', `SELECT_SCAN` bigint(20) unsigned NOT NULL COMMENT 'join時，第一個表位全表掃描的數目', `SORT_MERGE_PASSES` bigint(20) unsigned NOT NULL COMMENT '', `SORT_RANGE` bigint(20) unsigned NOT NULL COMMENT '范圍排序數目', `SORT_ROWS` bigint(20) unsigned NOT NULL COMMENT '排序的記錄數目', `SORT_SCAN` bigint(20) unsigned NOT NULL COMMENT '全表排序數目', `NO_INDEX_USED` bigint(20) unsigned NOT NULL COMMENT '沒有使用索引數目', `NO_GOOD_INDEX_USED` bigint(20) unsigned NOT NULL COMMENT '', `NESTING_EVENT_ID` bigint(20) unsigned DEFAULT NULL COMMENT '該事件對應的父事件ID', `NESTING_EVENT_TYPE` enum('STATEMENT','STAGE','WAIT') DEFAULT NULL COMMENT '父事件類型(STATEMENT, STAGE, WAIT)' ) ENGINE=PERFORMANCE_SCHEMA DEFAULT CHARSET=utf8
六：Connection 表
1，users：記錄用戶連接數信息
2，hosts：記錄了主機連接數信息
3，accounts：記錄了用戶主機連接數信息
zjy@performance_schema 12:03:27>select * from users; +------------------+---------------------+-------------------+ | USER | CURRENT_CONNECTIONS | TOTAL_CONNECTIONS | +------------------+---------------------+-------------------+ | debian-sys-maint | 0 | 36 | | zjy | 1 | 22285 | | dchat_php | 0 | 37864 | | dxyslave | 2 | 9 | | nagios | 0 | 10770 | | dchat_data | 140 | 2233023 | | NULL | 0 | 15866 | | dchat_api | 160 | 2754212 | | mha_data | 1 | 36 | | backup | 0 | 15 | | cacti | 0 | 4312 | | kol | 10 | 172414 | +------------------+---------------------+-------------------+ 12 rows in set (0.00 sec) zjy@performance_schema 12:03:34>select * from hosts; +-----------------+---------------------+-------------------+ | HOST | CURRENT_CONNECTIONS | TOTAL_CONNECTIONS | +-----------------+---------------------+-------------------+ | 192.168.100.218 | 150 | 2499422 | | 192.168.100.240 | 10 | 172429 | | 192.168.100.139 | 0 | 698 | | 192.168.100.21 | 0 | 2 | | 192.168.100.220 | 150 | 2526136 | | 192.168.100.25 | 1 | 7 | | NULL | 0 | 15867 | | 192.168.100.241 | 0 | 21558 | | 192.168.100.191 | 1 | 34 | | localhost | 0 | 10807 | | 192.168.100.118 | 1 | 2 | | 192.168.100.251 | 0 | 4312 | | 192.168.100.23 | 1 | 31 | | 192.168.100.193 | 0 | 15 | +-----------------+---------------------+-------------------+ 14 rows in set (0.01 sec) zjy@performance_schema 12:05:21>select * from accounts; +------------------+-----------------+---------------------+-------------------+ | USER | HOST | CURRENT_CONNECTIONS | TOTAL_CONNECTIONS | +------------------+-----------------+---------------------+-------------------+ | cacti | 192.168.100.251 | 0 | 4313 | | debian-sys-maint | localhost | 0 | 36 | | backup | 192.168.100.193 | 0 | 15 | | dchat_api | 192.168.100.220 | 80 | 1382585 | | dchat_php | 192.168.100.220 | 0 | 20292 | | zjy | 192.168.100.139 | 0 | 698 | | zjy | 192.168.100.241 | 0 | 21558 | | mha_data | 192.168.100.191 | 1 | 34 | | dxyslave | 192.168.100.118 | 1 | 2 | | kol | 192.168.100.240 | 10 | 172431 | | dxyslave | 192.168.100.25 | 1 | 7 | | dchat_data | 192.168.100.218 | 70 | 1109974 | | zjy | 192.168.100.23 | 1 | 31 | | dchat_php | 192.168.100.218 | 0 | 17572 | | dchat_data | 192.168.100.220 | 70 | 1123306 | | NULL | NULL | 0 | 15868 | | mha_data | 192.168.100.21 | 0 | 2 | | dchat_api | 192.168.100.218 | 80 | 1371918 | | nagios | localhost | 0 | 10771 | +------------------+-----------------+---------------------+-------------------+
七：Summary 表： Summary表聚集了各個維度的統計信息包括表維度，索引維度，會話維度，語句維度和鎖維度的統計信息
1，events_waits_summary_global_by_event_name：按等待事件類型聚合，每個事件一條記錄
CREATE TABLE `events_waits_summary_global_by_event_name` ( `EVENT_NAME` varchar(128) NOT NULL COMMENT '事件名稱', `COUNT_STAR` bigint(20) unsigned NOT NULL COMMENT '事件計數', `SUM_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '總的等待時間', `MIN_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最小等待時間', `AVG_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '平均等待時間', `MAX_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最大等待時間' ) ENGINE=PERFORMANCE_SCHEMA DEFAULT CHARSET=utf8
2，events_waits_summary_by_instance：按等待事件對象聚合，同一種等待事件，可能有多個實例，每個實例有不同的內存地址，因此
event_name+object_instance_begin唯一確定一條記錄。
CREATE TABLE `events_waits_summary_by_instance` ( `EVENT_NAME` varchar(128) NOT NULL COMMENT '事件名稱', `OBJECT_INSTANCE_BEGIN` bigint(20) unsigned NOT NULL COMMENT '內存地址', `COUNT_STAR` bigint(20) unsigned NOT NULL COMMENT '事件計數', `SUM_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '總的等待時間', `MIN_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最小等待時間', `AVG_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '平均等待時間', `MAX_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最大等待時間' ) ENGINE=PERFORMANCE_SCHEMA DEFAULT CHARSET=utf8
3，events_waits_summary_by_thread_by_event_name：按每個線程和事件來統計，thread_id+event_name唯一確定一條記錄。
CREATE TABLE `events_waits_summary_by_thread_by_event_name` ( `THREAD_ID` bigint(20) unsigned NOT NULL COMMENT '線程ID', `EVENT_NAME` varchar(128) NOT NULL COMMENT '事件名稱', `COUNT_STAR` bigint(20) unsigned NOT NULL COMMENT '事件計數', `SUM_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '總的等待時間', `MIN_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最小等待時間', `AVG_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '平均等待時間', `MAX_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最大等待時間' ) ENGINE=PERFORMANCE_SCHEMA DEFAULT CHARSET=utf8
4，events_stages_summary_global_by_event_name：按事件階段類型聚合，每個事件一條記錄，表結構同上。
5，events_stages_summary_by_thread_by_event_name：按每個線程和事件來階段統計，表結構同上。
6，events_statements_summary_by_digest：按照事件的語句進行聚合。
CREATE TABLE `events_statements_summary_by_digest` ( `SCHEMA_NAME` varchar(64) DEFAULT NULL COMMENT '庫名', `DIGEST` varchar(32) DEFAULT NULL COMMENT '對SQL_TEXT做MD5產生的32位字符串。如果為consumer表中沒有打開statement_digest選項，則為NULL', `DIGEST_TEXT` longtext COMMENT '將語句中值部分用問號代替，用於SQL語句歸類。如果為consumer表中沒有打開statement_digest選項，則為NULL。', `COUNT_STAR` bigint(20) unsigned NOT NULL COMMENT '事件計數', `SUM_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '總的等待時間', `MIN_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最小等待時間', `AVG_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '平均等待時間', `MAX_TIMER_WAIT` bigint(20) unsigned NOT NULL COMMENT '最大等待時間', `SUM_LOCK_TIME` bigint(20) unsigned NOT NULL COMMENT '鎖時間總時長', `SUM_ERRORS` bigint(20) unsigned NOT NULL COMMENT '錯誤數的總', `SUM_WARNINGS` bigint(20) unsigned NOT NULL COMMENT '警告的總數', `SUM_ROWS_AFFECTED` bigint(20) unsigned NOT NULL COMMENT '影響的總數目', `SUM_ROWS_SENT` bigint(20) unsigned NOT NULL COMMENT '返回總數目', `SUM_ROWS_EXAMINED` bigint(20) unsigned NOT NULL COMMENT '總的掃描的數目', `SUM_CREATED_TMP_DISK_TABLES` bigint(20) unsigned NOT NULL COMMENT '創建磁盤臨時表的總數目', `SUM_CREATED_TMP_TABLES` bigint(20) unsigned NOT NULL COMMENT '創建臨時表的總數目', `SUM_SELECT_FULL_JOIN` bigint(20) unsigned NOT NULL COMMENT '第一個表全表掃描的總數目', `SUM_SELECT_FULL_RANGE_JOIN` bigint(20) unsigned NOT NULL COMMENT '總的采用range方式掃描的數目', `SUM_SELECT_RANGE` bigint(20) unsigned NOT NULL COMMENT '第一個表采用range方式掃描的總數目', `SUM_SELECT_RANGE_CHECK` bigint(20) unsigned NOT NULL COMMENT '', `SUM_SELECT_SCAN` bigint(20) unsigned NOT NULL COMMENT '第一個表位全表掃描的總數目', `SUM_SORT_MERGE_PASSES` bigint(20) unsigned NOT NULL COMMENT '', `SUM_SORT_RANGE` bigint(20) unsigned NOT NULL COMMENT '范圍排序總數', `SUM_SORT_ROWS` bigint(20) unsigned NOT NULL COMMENT '排序的記錄總數目', `SUM_SORT_SCAN` bigint(20) unsigned NOT NULL COMMENT '第一個表排序掃描總數目', `SUM_NO_INDEX_USED` bigint(20) unsigned NOT NULL COMMENT '沒有使用索引總數', `SUM_NO_GOOD_INDEX_USED` bigint(20) unsigned NOT NULL COMMENT '', `FIRST_SEEN` timestamp NOT NULL DEFAULT '0000-00-00 00:00:00' COMMENT '第一次執行時間', `LAST_SEEN` timestamp NOT NULL DEFAULT '0000-00-00 00:00:00' COMMENT '最后一次執行時間' ) ENGINE=PERFORMANCE_SCHEMA DEFAULT CHARSET=utf8
7，events_statements_summary_global_by_event_name：按照事件的語句進行聚合。表結構同上。
8，events_statements_summary_by_thread_by_event_name：按照線程和事件的語句進行聚合，表結構同上。
9，file_summary_by_instance：按事件類型統計（物理IO維度）
10，file_summary_by_event_name：具體文件統計（物理IO維度）
9和10一起說明：
統計IO操作：COUNT_STAR，SUM_TIMER_WAIT,MIN_TIMER_WAIT,AVG_TIMER_WAIT,MAX_TIMER_WAIT
統計讀 ：COUNT_READ,SUM_TIMER_READ,MIN_TIMER_READ,AVG_TIMER_READ,MAX_TIMER_READ, SUM_NUMBER_OF_BYTES_READ
統計寫 ：COUNT_WRITE,SUM_TIMER_WRITE,MIN_TIMER_WRITE,AVG_TIMER_WRITE,MAX_TIMER_WRITE, SUM_NUMBER_OF_BYTES_WRITE
統計其他IO事件，比如create，delete，open，close等：COUNT_MISC,SUM_TIMER_MISC,MIN_TIMER_MISC,AVG_TIMER_MISC,MAX_TIMER_MISC
11，table_io_waits_summary_by_table：根據wait/io/table/sql/handler，聚合每個表的I/O操作（邏輯IO緯度）
統計IO操作：COUNT_STAR,SUM_TIMER_WAIT,MIN_TIMER_WAIT,AVG_TIMER_WAIT,MAX_TIMER_WAIT
統計讀 ：COUNT_READ,SUM_TIMER_READ,MIN_TIMER_READ,AVG_TIMER_READ,MAX_TIMER_READ
：COUNT_FETCH,SUM_TIMER_FETCH,MIN_TIMER_FETCH,AVG_TIMER_FETCH, MAX_TIMER_FETCH
統計寫 ：COUNT_WRITE,SUM_TIMER_WRITE,MIN_TIMER_WRITE,AVG_TIMER_WRITE,MAX_TIMER_WRITE
INSERT統計，相應的還有DELETE和UPDATE統計：COUNT_INSERT,SUM_TIMER_INSERT,MIN_TIMER_INSERT,AVG_TIMER_INSERT,MAX_TIMER_INSERT
12，table_io_waits_summary_by_index_usage：與table_io_waits_summary_by_table類似，按索引維度統計
13，table_lock_waits_summary_by_table：聚合了表鎖等待事件，包括internal lock 和 external lock
internal lock通過SQL層函數thr_lock調用，OPERATION值為：
read normal、read with shared locks、read high priority、read no insert、write allow write、write concurrent insert、write delayed、write low priority、write normal
external lock則通過接口函數handler::external_lock調用存儲引擎層，OPERATION列的值為：read external、write external
14，Connection Summaries表：account、user、host
events_waits_summary_by_account_by_event_name
events_waits_summary_by_user_by_event_name
events_waits_summary_by_host_by_event_name
events_stages_summary_by_account_by_event_name
events_stages_summary_by_user_by_event_name
events_stages_summary_by_host_by_event_name
events_statements_summary_by_account_by_event_name
events_statements_summary_by_user_by_event_name
events_statements_summary_by_host_by_event_name
15，socket_summary_by_instance、socket_summary_by_event_name：socket聚合統計表。
八：其他相關表
1，performance_timers：系統支持的統計時間單位
2，threads：監視服務端的當前運行的線程
統計應用：
關於SQL維度的統計信息主要集中在events_statements_summary_by_digest表中，通過將SQL語句抽象出digest，可以統計某類SQL語句在各個維度的統計信息
1，哪個SQL執行最多：
zjy@performance_schema 11:36:22>SELECT SCHEMA_NAME,DIGEST_TEXT,COUNT_STAR,SUM_ROWS_SENT,SUM_ROWS_EXAMINED,FIRST_SEEN,LAST_SEEN FROM events_statements_summary_by_digest ORDER BY COUNT_STAR desc LIMIT 1\G *************************** 1. row *************************** SCHEMA_NAME: dchat DIGEST_TEXT: SELECT ... COUNT_STAR: 1161210102 SUM_ROWS_SENT: 1161207842 SUM_ROWS_EXAMINED: 0 FIRST_SEEN: 2016-02-17 00:36:46 LAST_SEEN: 2016-03-07 11:36:29
各個字段的注釋可以看上面的表結構說明：從2月17號到3月7號該SQL執行了1161210102次。
2，哪個SQL平均響應時間最多：
zjy@performance_schema 11:36:28>SELECT SCHEMA_NAME,DIGEST_TEXT,COUNT_STAR,AVG_TIMER_WAIT,SUM_ROWS_SENT,SUM_ROWS_EXAMINED,FIRST_SEEN,LAST_SEEN FROM events_statements_summary_by_digest ORDER BY AVG_TIMER_WAIT desc LIMIT 1\G *************************** 1. row *************************** SCHEMA_NAME: dchat DIGEST_TEXT: SELECT ... COUNT_STAR: 1 AVG_TIMER_WAIT: 273238183964000 SUM_ROWS_SENT: 50208 SUM_ROWS_EXAMINED: 5565651 FIRST_SEEN: 2016-02-22 13:27:33 LAST_SEEN: 2016-02-22 13:27:33
各個字段的注釋可以看上面的表結構說明：從2月17號到3月7號該SQL平均響應時間273238183964000皮秒（1000000000000皮秒=1秒）
3，哪個SQL掃描的行數最多：
SUM_ROWS_EXAMINED
4，哪個SQL使用的臨時表最多：
SUM_CREATED_TMP_DISK_TABLES、SUM_CREATED_TMP_TABLES
5，哪個SQL返回的結果集最多：
SUM_ROWS_SENT
6，哪個SQL排序數最多：
SUM_SORT_ROWS
通過上述指標我們可以間接獲得某類SQL的邏輯IO(SUM_ROWS_EXAMINED)，CPU消耗(SUM_SORT_ROWS)，網絡帶寬(SUM_ROWS_SENT)的對比。
通過file_summary_by_instance表，可以獲得系統運行到現在，哪個文件(表)物理IO最多，這可能意味着這個表經常需要訪問磁盤IO。
7，哪個表、文件邏輯IO最多（熱數據）：
zjy@performance_schema 12:16:18>SELECT FILE_NAME,EVENT_NAME,COUNT_READ,SUM_NUMBER_OF_BYTES_READ,COUNT_WRITE,SUM_NUMBER_OF_BYTES_WRITE FROM file_summary_by_instance ORDER BY SUM_NUMBER_OF_BYTES_READ+SUM_NUMBER_OF_BYTES_WRITE DESC LIMIT 2\G *************************** 1. row *************************** FILE_NAME: /var/lib/mysql/ibdata1 #文件 EVENT_NAME: wait/io/file/innodb/innodb_data_file COUNT_READ: 544 SUM_NUMBER_OF_BYTES_READ: 10977280 COUNT_WRITE: 3700729 SUM_NUMBER_OF_BYTES_WRITE: 1433734217728 *************************** 2. row *************************** FILE_NAME: /var/lib/mysql/dchat/fans.ibd #表 EVENT_NAME: wait/io/file/innodb/innodb_data_file COUNT_READ: 9370680 SUM_NUMBER_OF_BYTES_READ: 153529188352 COUNT_WRITE: 67576376 SUM_NUMBER_OF_BYTES_WRITE: 1107815432192
8，哪個索引使用最多：
zjy@performance_schema 12:18:42>SELECT OBJECT_NAME, INDEX_NAME, COUNT_FETCH, COUNT_INSERT, COUNT_UPDATE, COUNT_DELETE FROM table_io_waits_summary_by_index_usage ORDER BY SUM_TIMER_WAIT DESC limit 1; +-------------+------------+-------------+--------------+--------------+--------------+ | OBJECT_NAME | INDEX_NAME | COUNT_FETCH | COUNT_INSERT | COUNT_UPDATE | COUNT_DELETE | +-------------+------------+-------------+--------------+--------------+--------------+ | fans | PRIMARY | 29002695158 | 0 | 296373434 | 0 | +-------------+------------+-------------+--------------+--------------+--------------+ 1 row in set (0.29 sec)
通過table_io_waits_summary_by_index_usage表，可以獲得系統運行到現在，哪個表的具體哪個索引(包括主鍵索引，二級索引)使用最多。
9，哪個索引沒有使用過：
zjy@performance_schema 12:23:22>SELECT OBJECT_SCHEMA, OBJECT_NAME, INDEX_NAME FROM table_io_waits_summary_by_index_usage WHERE INDEX_NAME IS NOT NULL AND COUNT_STAR = 0 AND OBJECT_SCHEMA <> 'mysql' ORDER BY OBJECT_SCHEMA,OBJECT_NAME;
10，哪個等待事件消耗的時間最多：
zjy@performance_schema 12:25:22>SELECT EVENT_NAME, COUNT_STAR, SUM_TIMER_WAIT, AVG_TIMER_WAIT FROM events_waits_summary_global_by_event_name WHERE event_name != 'idle' ORDER BY SUM_TIMER_WAIT DESC LIMIT 1;
11，類似profiling功能：
分析具體某條SQL，該SQL在執行各個階段的時間消耗，通過events_statements_xxx表和events_stages_xxx表，就可以達到目的。兩個表通過event_id與nesting_event_id關聯，stages表的nesting_event_id為對應statements表的event_id；針對每個stage可能出現的鎖等待，一個stage會對應一個或多個wait，通過stage_xxx表的event_id字段與waits_xxx表的nesting_event_id進行關聯。如：
比如分析包含count(*)的某條SQL語句，具體如下： SELECT EVENT_ID, sql_text FROM events_statements_history WHERE sql_text LIKE '%count(*)%'; +----------+--------------------------------------+ | EVENT_ID | sql_text | +----------+--------------------------------------+ | 1690 | select count(*) from chuck.test_slow | +----------+--------------------------------------+ 首先得到了語句的event_id為1690，通過查找events_stages_xxx中nesting_event_id為1690的記錄，可以達到目的。 a.查看每個階段的時間消耗： SELECT event_id, EVENT_NAME, SOURCE, TIMER_END - TIMER_START FROM events_stages_history_long WHERE NESTING_EVENT_ID = 1690; +----------+--------------------------------+----------------------+-----------------------+ | event_id | EVENT_NAME | SOURCE | TIMER_END-TIMER_START | +----------+--------------------------------+----------------------+-----------------------+ | 1691 | stage/sql/init | mysqld.cc:990 | 316945000 | | 1693 | stage/sql/checking permissions | sql_parse.cc:5776 | 26774000 | | 1695 | stage/sql/Opening tables | sql_base.cc:4970 | 41436934000 | | 2638 | stage/sql/init | sql_select.cc:1050 | 85757000 | | 2639 | stage/sql/System lock | lock.cc:303 | 40017000 | | 2643 | stage/sql/optimizing | sql_optimizer.cc:138 | 38562000 | | 2644 | stage/sql/statistics | sql_optimizer.cc:362 | 52845000 | | 2645 | stage/sql/preparing | sql_optimizer.cc:485 | 53196000 | | 2646 | stage/sql/executing | sql_executor.cc:112 | 3153000 | | 2647 | stage/sql/Sending data | sql_executor.cc:192 | 7369072089000 | | 4304138 | stage/sql/end | sql_select.cc:1105 | 19920000 | | 4304139 | stage/sql/query end | sql_parse.cc:5463 | 44721000 | | 4304145 | stage/sql/closing tables | sql_parse.cc:5524 | 61723000 | | 4304152 | stage/sql/freeing items | sql_parse.cc:6838 | 455678000 | | 4304155 | stage/sql/logging slow query | sql_parse.cc:2258 | 83348000 | | 4304159 | stage/sql/cleaning up | sql_parse.cc:2163 | 4433000 | +----------+--------------------------------+----------------------+-----------------------+ 通過間接關聯，我們能分析得到SQL語句在每個階段的時間消耗，時間單位以皮秒表示。這里展示的結果很類似profiling功能，有了performance schema，就不再需要profiling這個功能了。另外需要注意的是，由於默認情況下events_stages_history表中只為每個連接記錄了最近10條記錄，為了確保獲取所有記錄，需要訪問events_stages_history_long表 b.查看某個階段的鎖等待情況 針對每個stage可能出現的鎖等待，一個stage會對應一個或多個wait，events_waits_history_long這個表容易爆滿[默認閥值10000]。由於select count(*)需要IO(邏輯IO或者物理IO)，所以在stage/sql/Sending data階段會有io等待的統計。通過stage_xxx表的event_id字段與waits_xxx表的nesting_event_id進行關聯。 SELECT event_id, event_name, source, timer_wait, object_name, index_name, operation, nesting_event_id FROM events_waits_history_long WHERE nesting_event_id = 2647; +----------+---------------------------+-----------------+------------+-------------+------------+-----------+------------------+ | event_id | event_name | source | timer_wait | object_name | index_name | operation | nesting_event_id | +----------+---------------------------+-----------------+------------+-------------+------------+-----------+------------------+ | 190607 | wait/io/table/sql/handler | handler.cc:2842 | 1845888 | test_slow | idx_c1 | fetch | 2647 | | 190608 | wait/io/table/sql/handler | handler.cc:2842 | 1955328 | test_slow | idx_c1 | fetch | 2647 | | 190609 | wait/io/table/sql/handler | handler.cc:2842 | 1929792 | test_slow | idx_c1 | fetch | 2647 | | 190610 | wait/io/table/sql/handler | handler.cc:2842 | 1869600 | test_slow | idx_c1 | fetch | 2647 | | 190611 | wait/io/table/sql/handler | handler.cc:2842 | 1922496 | test_slow | idx_c1 | fetch | 2647 | +----------+---------------------------+-----------------+------------+-------------+------------+-----------+------------------+ 通過上面的實驗，我們知道了statement,stage,wait的三級結構，通過nesting_event_id進行關聯，它表示某個事件的父event_id。 (2).模擬innodb行鎖等待的例子 會話A執行語句update test_icp set y=y+1 where x=1(x為primary key)，不commit；會話B執行同樣的語句update test_icp set y=y+1 where x=1，會話B堵塞，並最終報錯。通過連接連接查詢events_statements_history_long和events_stages_history_long，可以看到在updating階段花了大約60s的時間。這主要因為實例上的innodb_lock_wait_timeout設置為60，等待60s后超時報錯了。 SELECT statement.EVENT_ID, stages.event_id, statement.sql_text, stages.event_name, stages.timer_wait FROM events_statements_history_long statement join events_stages_history_long stages on statement.event_id=stages.nesting_event_id WHERE statement.sql_text = 'update test_icp set y=y+1 where x=1'; +----------+----------+-------------------------------------+--------------------------------+----------------+ | EVENT_ID | event_id | sql_text | event_name | timer_wait | +----------+----------+-------------------------------------+--------------------------------+----------------+ | 5816 | 5817 | update test_icp set y=y+1 where x=1 | stage/sql/init | 195543000 | | 5816 | 5819 | update test_icp set y=y+1 where x=1 | stage/sql/checking permissions | 22730000 | | 5816 | 5821 | update test_icp set y=y+1 where x=1 | stage/sql/Opening tables | 66079000 | | 5816 | 5827 | update test_icp set y=y+1 where x=1 | stage/sql/init | 89116000 | | 5816 | 5828 | update test_icp set y=y+1 where x=1 | stage/sql/System lock | 218744000 | | 5816 | 5832 | update test_icp set y=y+1 where x=1 | stage/sql/updating | 6001362045000 | | 5816 | 5968 | update test_icp set y=y+1 where x=1 | stage/sql/end | 10435000 | | 5816 | 5969 | update test_icp set y=y+1 where x=1 | stage/sql/query end | 85979000 | | 5816 | 5983 | update test_icp set y=y+1 where x=1 | stage/sql/closing tables | 56562000 | | 5816 | 5990 | update test_icp set y=y+1 where x=1 | stage/sql/freeing items | 83563000 | | 5816 | 5992 | update test_icp set y=y+1 where x=1 | stage/sql/cleaning up | 4589000 | +----------+----------+-------------------------------------+--------------------------------+----------------+ 查看wait事件： SELECT event_id, event_name, source, timer_wait, object_name, index_name, operation, nesting_event_id FROM events_waits_history_long WHERE nesting_event_id = 5832; *************************** 1. row *************************** event_id: 5832 event_name: wait/io/table/sql/handler source: handler.cc:2782 timer_wait: 6005946156624 object_name: test_icp index_name: PRIMARY operation: fetch 從結果來看，waits表中記錄了一個fetch等待事件，但並沒有更細的innodb行鎖等待事件統計。 (3).模擬MDL鎖等待的例子 會話A執行一個大查詢select count(*) from test_slow，會話B執行表結構變更alter table test_slow modify c2 varchar(152);通過如下語句可以得到alter語句的執行過程，重點關注“stage/sql/Waiting for table metadata lock”階段。 SELECT statement.EVENT_ID, stages.event_id, statement.sql_text, stages.event_name as stage_name, stages.timer_wait as stage_time FROM events_statements_history_long statement left join events_stages_history_long stages on statement.event_id=stages.nesting_event_id WHERE statement.sql_text = 'alter table test_slow modify c2 varchar(152)'; +-----------+-----------+----------------------------------------------+----------------------------------------------------+---------------+ | EVENT_ID | event_id | sql_text | stage_name | stage_time | +-----------+-----------+----------------------------------------------+----------------------------------------------------+---------------+ | 326526744 | 326526745 | alter table test_slow modify c2 varchar(152) | stage/sql/init | 216662000 | | 326526744 | 326526747 | alter table test_slow modify c2 varchar(152) | stage/sql/checking permissions | 18183000 | | 326526744 | 326526748 | alter table test_slow modify c2 varchar(152) | stage/sql/checking permissions | 10294000 | | 326526744 | 326526750 | alter table test_slow modify c2 varchar(152) | stage/sql/init | 4783000 | | 326526744 | 326526751 | alter table test_slow modify c2 varchar(152) | stage/sql/Opening tables | 140172000 | | 326526744 | 326526760 | alter table test_slow modify c2 varchar(152) | stage/sql/setup | 157643000 | | 326526744 | 326526769 | alter table test_slow modify c2 varchar(152) | stage/sql/creating table | 8723217000 | | 326526744 | 326526803 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 257332000 | | 326526744 | 326526832 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 1000181831000 | | 326526744 | 326526835 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 33483000 | | 326526744 | 326526838 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 1000091810000 | | 326526744 | 326526841 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 17187000 | | 326526744 | 326526844 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 1000126464000 | | 326526744 | 326526847 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 27472000 | | 326526744 | 326526850 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 561996133000 | | 326526744 | 326526853 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 124876000 | | 326526744 | 326526877 | alter table test_slow modify c2 varchar(152) | stage/sql/System lock | 30659000 | | 326526744 | 326526881 | alter table test_slow modify c2 varchar(152) | stage/sql/preparing for alter table | 40246000 | | 326526744 | 326526889 | alter table test_slow modify c2 varchar(152) | stage/sql/altering table | 36628000 | | 326526744 | 326528280 | alter table test_slow modify c2 varchar(152) | stage/sql/end | 43824000 | | 326526744 | 326528281 | alter table test_slow modify c2 varchar(152) | stage/sql/query end | 112557000 | | 326526744 | 326528299 | alter table test_slow modify c2 varchar(152) | stage/sql/closing tables | 27707000 | | 326526744 | 326528305 | alter table test_slow modify c2 varchar(152) | stage/sql/freeing items | 201614000 | | 326526744 | 326528308 | alter table test_slow modify c2 varchar(152) | stage/sql/cleaning up | 3584000 | +-----------+-----------+----------------------------------------------+----------------------------------------------------+---------------+ 從結果可以看到，出現了多次stage/sql/Waiting for table metadata lock階段，並且間隔1s，說明每隔1s鍾會重試判斷。找一個該階段的event_id,通過nesting_event_id關聯，確定到底在等待哪個wait事件。 SELECT event_id, event_name, source, timer_wait, object_name, index_name, operation, nesting_event_id FROM events_waits_history_long WHERE nesting_event_id = 326526850; +-----------+---------------------------------------------------+------------------+--------------+-------------+------------+------------+------------------+ | event_id | event_name | source | timer_wait | object_name | index_name | operation | nesting_event_id | +-----------+---------------------------------------------------+------------------+--------------+-------------+------------+------------+------------------+ | 326526851 | wait/synch/cond/sql/MDL_context::COND_wait_status | mdl.cc:1327 | 562417991328 | NULL | NULL | timed_wait | 326526850 | | 326526852 | wait/synch/mutex/mysys/my_thread_var::mutex | sql_class.h:3481 | 733248 | NULL | NULL | lock | 326526850 | +-----------+---------------------------------------------------+------------------+--------------+-------------+------------+------------+------------------+ 通過結果可以知道，產生阻塞的是條件變量MDL_context::COND_wait_status，並且顯示了代碼的位置。
總結：
本文通過對Performance Schema數據庫的介紹，主要用於收集數據庫服務器性能參數：①提供進程等待的詳細信息，包括鎖、互斥變量、文件信息；②保存歷史的事件匯總信息，為提供MySQL服務器性能做出詳細的判斷；③對於新增和刪除監控事件點都非常容易，並可以改變mysql服務器的監控周期，例如（CYCLE、MICROSECOND）。通過該庫得到數據庫運行的統計信息，更好分析定位問題和完善監控信息。類似的監控還有：
打開標准的innodb監控： CREATE TABLE innodb_monitor (a INT) ENGINE=INNODB; 打開innodb的鎖監控： CREATE TABLE innodb_lock_monitor (a INT) ENGINE=INNODB; 打開innodb表空間監控： CREATE TABLE innodb_tablespace_monitor (a INT) ENGINE=INNODB; 打開innodb表監控： CREATE TABLE innodb_table_monitor (a INT) ENGINE=INNODB;
參考文章：
https://dev.mysql.com/doc/refman/5.6/en/performance-schema.html
http://www.cnblogs.com/cchust/p/5022148.html
http://www.cnblogs.com/cchust/p/5057498.html
http://www.cnblogs.com/cchust/p/5061131.html