A DB2 Performance Tuning Roadmap --DIVE INTO LOCK

在整理了DB2 LOG相关内容的基础之上，这章整理lock的内容，相比较与log，lock的内容更多的与应用相关，涉及内容方方面面更加复杂。DBMS在本质上不同于文件系统的地方在于DBMS系统所支持的事务。 如果log的引入是为了保证持久化，那么LOCK的引入就是保证事务串行化，解决事务的并发带来了资源的RACE CONDITION。不同的事务场景，定义了不同的并发需求，为此定义了4种事务的隔离级别，不同的DBMS引入了不同的锁机制来实现。DB2的串行化机制主要是通过LOCK,LATCH,CLAIM/DRAIM 来实现，具体的LOCK的相关属性以及LOCK属性对应用程序的影响比如OBJECT,SIZE,MODE,DURATION,PARTICIPANTS,PARAMETER LOCATIONS都进行了介绍。当引入了DBMS CLUSTER(DB2中为DATA SHARING GROUP,而ORACLE中为ORACEL RAC) 以后，为了处理不同不同MEMBER之间数据的一致性，DB2引入了PHYSICAL LOCK以及CF LOCK STRTURE 来实现全局LOCK 冲突检测。单个SUBSYSTEM中，主要的冲突有TIMEOUT,DEADLOCK,在DATASHARING GROUP,引入了新的CONTENTION,XES CONTENTION,FALSE CONTENTION,GLOBAL LOCK CONTENTION。引入LOCK的同时不可避免的带来了新的OVERHEAD，DB2主要通过IRLM,XES,CF等处理锁资源的请求，除了地址空间除了锁请求之外，由于应用本身设计不合理或是某些特定的场景带来了很多问题，如TEMEOUT,DEADLOCK,LOCK ESCALATION等，如何有效的避免这些问题，需要系统运维人员以及开发人员共同努力。
本文的行文脉络基于个人对DBMS LOCK的认知层次，行文的逻辑性，合理性,整理内容的知识面的广度和深度都有待进一步的思考。
这篇博客从落笔到完成大体的框架，持持续续时间已经接近2周，应该是自己耗时最长的一篇博客，纵使如此，每一次查看，发下仍有新的东西需要自己补充，等后续会继续补充自己的理解。

• LOCK OVERVIEW
  • WHY LOCKS ? DB2 Serialization Mechanisms
    • Data Consistency And Database Concurrency
      • Phenomena Seen When Transactions Run Concurrently
      • Lost Updates
      • DIRTY READS
      • Non-Repeatable Reads
      • Phantoms
  • LOCK PROPERTIES
    • LOCK OBJECT OWNER
    • LOCK PARTICIPANTS
    • LOCK SIZE
    • LOCK MODE
    • LOCK DURATION
    • LOCK REQUEST AND RELEASE
    • LOCK AND TRAN ISOLATION LELVEL
    • LOCKING PARAMETERS LOCATIONS
      • DDL
      • DML
      • Precompiler Locking Parameters
      • Bind Locking Parameters
      • Zparm Locking Parameters
  • CLAIM AND DRAIN
  • LATCH
  • ADVANCED TOPICS
    • L-LOCK(EXPLICIT HIERARCHICAL LOCKING )EHL
    • PHYSICAL LOCK
    • RETAINED LOCKS(UPDATE LOCKS)
      • IMPACT OF Retained LOCKS
    • LOCK SCOPE:
      • LOCAL LOCK
      • GLOBAL LOCK
      • GLOBAL LOCK COMMUNICATION
        • Page Set P-Lock Negotiation
    • CF LOCK STRUCTURE
      • MODIFIED RESOURCE LIST(MRL)
      • LOCK TABLE
      • XCF
      • XES
        • XES CONTENTION
        • FALSE CONTENTION
        • GLOBAL CONTENTION
        • DATA SHARING ACTIVITY REPORT
      • Lock Structure Shortage Actions
  • Types of locking problems
    • How do I find out I have a problem
    • Analyzing concurrency problems --TOOLBOXS
      • DB2 commands and EXPLAIN
      • DB2 TRACES
    • EXAMPLE:Analysis of a simple deadlock scenario and solution
      • DBD is locked

LOCK OVERVIEW

WHY LOCKS ?DB2 Serialization Mechanisms

Data Consistency And Database Concurrency

DBMS区别与文件系统的最本质区别：DBMS支持事务
DBMS :Allowing multiple users to access a database simultaneously without compromising data integrity.
A transaction (or unit of work)is a recoverable sequence of one or more SQL operations that are grouped together as a single unit, usually within an application process.
One of the mechanisms DB2 uses to keep data consistent is the transaction. A transaction or(otherwise known as a unit of work) is a recoverable sequence of one or more SQL operations that are grouped together as a single unit, usually within an application process. The initiation and termination of a single transaction defines points of data consistency within a database; either the effects of all SQL operations performed within a transaction are applied to the database and made permanent (committed), or the effects of all SQL operations performed are completely "undone" and thrown away (rolled back).

Phenomena Seen When Transactions Run Concurrently

事务并发带来的问题，为了解决这些问题，DB2定义了4种事务隔离级别。

1. LOST UPDATE
2. DIRTY READS
3. NON-REPEATABLE READS
4. PHANTOMS

Lost Updates

Occurs when two transactions read the same data, both attempt to update the data read, and one of the updates is lost

Transaction 1 and Transaction 2 read the same row of data and both calculate new values for that row based upon the original values read. If Transaction 1 updates the row with its new value and Transaction 2 then updates the same row, the update operation performed by Transaction 1 is lost 

DIRTY READS

Occurs when a transaction reads data that has not yet been committed

Transaction 1 changes a row of data and Transaction 2 reads the changed row before Transaction 1 commits the change. If Transaction 1 rolls back the change, Transaction 2 will have read data that theoretically, never existed. 

Non-Repeatable Reads

Occurs when a transaction executes the same query multiple times and gets different results with each execution

Transaction 1 reads a row of data, then Transaction 2 modifies or deletes that row and commits the change. When Transaction 1 attempts to reread the row, it will retrieve different data values 

Phantoms

Occurs when a row of data that matches some search criteria is not seen initially

Transaction 1 retrieves a set of rows that satisfy some search criteria, then Transaction 2 inserts a new row that contains matching search criteria for Transaction 1’s query. If Transaction 1 re-executes the query that produced the original set of rows, a different set of rows will be retrieved – the new row added by Transaction 2 will now be included in the set of rows returned 

LOCK PROPERTIES

LOCK OBJECT OWNER

这里的object是指锁所施加的对象，不同的锁所能适用的对象是不同的。如PAGE LOCK的对象肯定是PAGE.
OWNER 表示谁持有锁，有的是TRAN,有的是DB2 MEMBER
LOCK PARTICIPANTS

Locking is a complex interaction of many parts.

1. DBM1, where SQL executes, is the beginning of locking.
2. The IRLM, a separate address space, is where locks are held and managed. The DBM1 AS requests locks from the IRLM. In addition to its functionality of granting and releasing locks, the IRLM is also in charge of detecting deadlock and timeout situations.
3. In a data sharing environment, there are two other ingredients in the soup. The XCF address space is where its XCF component resides. The XES does some lock management and communicates directly with the lock structure in the coupling facility

   LOCK SIZE

   访问数据的范围
   设计锁的颗粒度对系统和应用均有影响：设计时要综合考虑
   1. 锁本身也是一种资源，持有的lock size越大，锁资源消耗越小
   2. lock size的大小与对应用访问的并发（concurrent）成反比关系
   3. LOCK SIZE 带来的另外一个副作用就是当你申请的锁资源足够多，以至于超过系统设定的阀值时（NUMLKUS/NUMLKTS/LOCK MAX =0,SYSTEM,N），
      lock size： PAGE/ROW /TABLE /PARENT LOCKS:TABLESPACE /PARTITION

      Page/Row locks are not compatible with tablespace locks The solution is "Intent Locking" Regardless of the locksize, DB2 will always start with a tablespace lock 

      INTENT LOCKS:IS/IX USED whenever page or row locks are being used
      GROSS LOCKS:The gross locks (S, U, SIX and X) are tablespace locks which are used in three
      situations
   4. lock size tablespace
   5. lock table for a non-segmented tablespace
   6. lock escalation for a non-partitioned tablespace

      LOCK MODE

      数据的访问方式：独占还是共享 S/U/X
      X:exclusive, not sharable with other
      S:shared, sharable with other S-locks
      LOCK MODE之间的兼容性如下：

      S	U	X
      S	Y	Y	N
      U	Y	N	N
      X	N	N	N

      LOCK DURATION

持有数据的时间：commit，across commit(with hold)
FROM:START OF FIRST USE
TO :COMMIT OR MOMENTIALIY

LOCK REQUEST AND RELEASE

When modifying data (Insert, Update and Delete),DB2 determines all locking actions

1. Lock size Tablespace
   => lock is taken at first use
   => lock is released at commit or deallocate
2. SQL- Statement “Lock Table … in Exclusive Mode”
   => initially an IX-lock (on the tablespace)
   => lock is taken at first use
   => lock is released at commit or deallocate
3. Page lock and row lock
   => initially an IX-lock (on the tablespace)
   => then X-locks are taken on each page/row as needed
   => these X-locks are released at commit

LOCK AND TRAN ISOLATION LELVEL

每一种事务隔离级别锁所能解决的问题

1. Repeatable Read
   => locks all data read by DB2
   => guarantees that exact the same result set will be returned when reusing the cursor
2. Read Stability
   => locks all data “seen” by the application
   => guarantees that at least the same result set will be returned when reusing the cursor
3. Uncommitted Read (UR)
   Pro:
   Our read will be “cheaper”
   We are not disturbing others
   Con:
   We may read data
   which may be “un-inserted”
   which may be “un-updated”
   When you are reading live (operational) data,you can never guarantee exact results
4. Cursor Stability (CS)
   => a lock is taken when the row is fetched
   => the lock is released when next row is fetched,
   or at end of data or at Close Cursor
   For page locks:
   => a lock is taken when the first row on the page is fetched
   => the lock is released when next page is fetched,
   or at end of data or at Close Cursor
   This means:When you have fetched a row and have not reached end of data or closed the cursor,
   you are holding an S-lock on the row or page
5. LOCK AVOIDANCE
   Updateable

   DECLARE upd_cur CURSOR FOR SELECT data1, data2 FROM table FOR UPDATE OF colx 

   Read-Only

   DECLARE ro_cur CURSOR FOR SELECT DEPTNO, AVG(SALARY) FROM EMP GROUP BY DEPTNO 

   Ambiguous

   DECLARE amb_cur CURSOR FOR SELECT data1, data2 FROM table 

   => Use FOR READ ONLY or FOR FETCH ONLY
   If we have the following:
   
   Read-Only (or Ambiguous) Cursor
   Cursor Stability
   Currentdata No
   it is possible that we can get "Lock Avoidance“
   Lock avoidance is advantageous!
   A possible problem with Currentdata No:
   Fetch a row
   ...
   Update (or Delete) Where Key = :hostkey
   can give SQLSTATE="02000" (Not Found)

LOCKING PARAMETERS LOCATIONS

下面分别介绍了系统参数，应用参数如何影响LOCK行为

DDL

OBJECT DEFINITION,CREATE TABLESPACE:

1. LOCKMAX (INTEGER,SYSTEM)
2. LOCKSIZE
   DEFAULT:ANY
   OPTION:ANY,TABLESPACE,TABLE,PAGE,ROW,LOB
3. MAXROW: INTEGER
   DEFAULT:255
   USED TO increase concurrency
   MAXROWS 1Used to emulate row-level locking
   locking without the costs of page p-lock processing
4. MEMBERCLUSTER:(P-LOCK)
   Clustering per member
   Reduce p-lock contentions forspace mappages
   Destroys clustering according to clustering index
5. TRACKMOD (P-LOCK)
   DEFAUTL:NO
   OPTION:YES/NO

DML

1. LOCK TABLE
   SYNTAX:LOCK {TABLE, TABLESPACE [partno]} IN
   {EXCLUSIVE, SHARE} MODE
2. Cursor Definitions
   FOR UPDATE OF
   FOR READ ONLY
   1. Tells DB2 the result set is read-only
   2. Positioned UPDATEs and DELETEs are not allowed
   3. No U or X locks will be obtained for the cursor
3. Isolation Clause
   WITH [UR, CS, RS, RR]
   1. USE AND KEEP EXCLUSIVE LOCKS
   2. USE AND KEEP UPDATE LOCKS
   3. USE AND KEEP SHARE LOCKS

Precompiler Locking Parameters

NOFOR Option:NO FOR update clause mandatory for positioned updates
STDSQL(YES): Implies NOFOR

Bind Locking Parameters

1. CURRENTDATA
   FUNCTION:CURRENTDATA helps to determine if DB2 will attempt to avoid locks
   mean:Must the data in program be equal to the data at the current cursor position
   lock aovidance
   demo:
   
   CURRENTDATA(NO)
   note
2. ISOLATION
3. RELEASE

Zparm Locking Parameters

CLAIM AND DRAIN

BETWEEN SQL AND SQL:DB2 USE LOCKS
BETWEEN UTILITY AND SQL:DB2 USE CLAIM AND DRAIN
BASE:
FOR EACH OBJECT(TABLESPACE,PARTITION,TABLE,IDNEXSPACE),THERE IS A CLAIM-COUNT
WHICH IS INCREASED 1 AT START AND DECREASE BY 1 AT COMMIT
A utility starts with a Drain
No new Claims are allowed (wait, maybe timeout)
If the Claim-count >0 the utility waits (maybe timeout)
When the Claim-count reaches zero, the utility can continue
Two important situations:

• UR (Uncommitted Read) takes no locks – but increases Claim-count (UR REQUEST MASS-DELETE LOCKS)
• For Cursor With Hold the Claim-count is not reduced at Commit

  A claim is a notification to DB2 that a particular object is currently being accessed. Claims usually do not continue to exist beyond the commit point one exception being a Cursor with Hold. In order to access the DB2 object within the next unit of work, an application needs to make a new claim.Claims notify DB2 that there is current interest in or activity on a DB2 object.Even if Uncommitted Read doesn’t take any locks, a claim is taken. As long as there are any claims on a DB2 object, no drains may be taken on the object until those claims are released. 

A drain is the action of obtaining access to a DB2 object, by:Preventing any new claims against the object.Waiting for all existing claims on the object to be released. 

A drain on a DB2 object causes DB2 to quiesce all applications currently claiming that resource, by allowing them to reach a commit point but preventing them (or any other application process) from making a new claim.A drain lock also prevents conflicting processes from trying to drain the same object at the same time.
Utilities detect claimers are present and wait
Drain Write waits for all write claims to be released
Drain All waits for claims on all classes to be released
SHRLEVEL(CHANGE) Utilities are CLAIMers

LATCH

Latches – managed by DB2
? BM page latching for index and data pages
? DB2 internal latching (many latches grouped into 32 latch classes)
? Latches – managed by IRLM
? Internal IRLM serialization

ADVANCED TOPICS

虽然标题是advanced topics,其实涉及的内容并没有本质的区别，只不过我们在看待问题，分析问题时需要从单个DB2 SUBSYSTEM的视角上升到DBMS CLUSTER的高度，即DATA SHARING GROUP的高度。为了处理不同的member之间共享数据的COHERENCY问题，DB2引入P-LOCK,区别主要在于单个DB2 SUBSYSTEM中的LOCK owner均为tran，而P-LOCK的owner为DB2 SUBSYSTEM.因此这里的内容按照这种分类进行介绍L-LOCK,P-LOCK,RETAINED LOCK，这几种锁之间的区别需要了解。

L-LOCK(EXPLICIT HIERARCHICAL LOCKING )EHL

特点：

THE KIND YOU HAVE IN BOTH DATA SHARING AND NON DATA SHARING SUBSYTEM THEN CONTROL DATA CONCRRRENCY OF ACCESS TO OBJECTS THEN CAN BE LOCAL OR GLOBAL THEY ASSOTIATED WITH PROGRAMS 

WHY IS HIERACHICAL

SO THE MEMBER DB2 SUBSYSTEM WILL NOT PROPAGATE L-LOCKS TO THE CF UNNECESSARILY 

1. PARENT LOCK
   PARENT LOCK ARE TABLE SPACE LOCKS OR DATA PARTITION LOCK
   GROSS LOCKS ARE ALWAYS GLOBAL LOCKS
2. CHILD LOCK
   PAGE /ROW LOCK
   EXCEPTION: MAYBE TABLES LOCKS IF SEGMENT TABLESPACE
   PAGE/ROW LOCK CAN BE GRANTED LOCALLY IF NOT IN INTER-DB2 R/W SHARING STATE
   ASYNCHRONOUS CHILD PROPAGATION WHEN INTER-SYTEM INTEREST FIST OCCUR
3. L-LOCK PROPAGATION
   NOT ALL L-LOCKS ARE PROPAGATED TO THE DB2 LOCK STRUCTURE
   PARENT L-LOCKS ARE ALMOST ALWAYS PROPAGATE TO THE DB2 LOCK STRUCTURE
4. Child Locks Propagation based on Pageset P-lock
   Now based on cached(held) state of the pageset P-lock
   If pageset P-lock negotiated from X to SIX or IX, then child L-locks propagated
   Reduced volatility
   If P-lock not held at time of child L-lock request, child lock will be propagated
   "Index-only" scan (if any locks taken) must open table space
   Parent L-lock no longer need to be held in cached state after DB2 failure
   A pageset IX L-lock no longer held as a retained X lock
   Important availability benefit in data sharing
   CHILD L-LOCKS ARE PROPAGATED TO THE DB2 LOCK STRUCTURE ONLY IF THERE IS INTER DB2 PARENT L-LOCKS CONFICT ON A DATABASES OBJECT

PHYSICAL LOCK

特点：
EXCLUSIVE TO DATA SHARING
ARE ONLY GLOBAL
USED TO MAINTAIN DATA COHERENCY IN DATA SHARING
ALSO USED TO MAINTAIN EDMPOOL CONSISTENCY AMONG MEMBERS
ASSOTIATED BY(OWNED BY) DB2 MEMBERS
P-LOCK CAN BE NEGOTIATTED BETWEEN DB2 MEMBERS
NO TIMEOUT OR DEADLOCK DETECTION
分类

1. PAGE SET P-LOCK
2. PAGE P-LOCK
3. PAGE SET/PARTITION CASTOUT P-LOCK

4. 
   

5. 
   

6. GBP STRUCTURE P-LOCK

   RETAINED LOCKS(UPDATE LOCKS)

   AIM:PROTECT UNCOMMITED DATA FROM ACCESS BY THER LOCK STRUCTURE
   UPDATE(IX,SIX,X)
   GLOBAL LOCKS ARE RETAINED AFTER FAILTURE

   IMPACT OF Retained LOCKS

   

   The ONLY way to clear retained locks is by estarting the failed DB2 subsystem. 

   LOCK SCOPE:

   LOCAL LOCK

GLOBAL LOCK

A LOCK THAT A DB2 SHARING DB2 MEMBERHAS TOMAKE KNOWN TO OTHERS MEMBERS OF THIS DATA SHARING GROUP
THE IRLM VIA XES,PROPAGATE TO THES LOCKS TO THE STRUCTURE IN THE CF
XCF,XES的内容在XES contention部分介绍。

GLOBAL LOCK COMMUNICATION

HOW ARE GLOBAL LOCK REQUESTS MADE KNOWN TO OTHER DB2 MEMBERS?

THEY ARE PROPAGATED TO DB2 LOCK STRUCTURE 

Page Set P-Lock Negotiation

与APPLICATION LOCK不同的是，PHYSCIAL LOCK 可以做Page Set P-Lock Negotiation。

上图展示了INERDB2 READ/WRITE INTEREST场景下的PAGE SET P-LOCK NEGOTIATION。所谓的negotiation就是锁请求方以及锁申请方将锁请求均DOWNGRADE，从而达到compatible的目的。需要注意的是，PAGE SET P-LOCK STATE状态的变化需要涉及IRLM以及XCF通讯机制。PAGE SET P-LOCK对GLOBAL BUFFER POOL起一种标志作用，
对于READ的DB2 SUBSYSTEM, must register its
interest in read pages in the GBP directory
对于UPDATER的DB2 SUBSYSTEM,must write updated pages to the group buffer pool and register those updated pages in the GBP directory

CF LOCK STRUCTURE

DEVIDE INTO TWO PARTS OF CF LOCK STRUCTURE RATIO OF LAST TWO COMPONETS DEFAULT ROUGHLY 50%CAN BE ALTERED BY IRLM PARAMETER 

MODIFIED RESOURCE LIST(MRL)

OVERVIEW OF MRL STRUCTURE

ACTIVEstatus of active means that the DB2 system is running.
Retained, on the other hand, means that a DB2 and/or its associated IRLM has failed
CONTAINS ENTRIES FOR ALL CRURENTLY HELD MODIFY LOCKS
IN THE EVENT OF DB2 FAILTURE ENSURE THAT OTHER MEMBER OF DB2 KNOW WHICH LOCKS MUST BERETIANED
It is essentially a list structure that lists all the update or modify locks in the data sharing group
DISPLAY GROUP ; Display “LIST ENTRIES IN USE”

 NUMBER LIST ENTRIES : 7353, LIST ENTRIES IN USE: 1834 

WARNING: When the LIST ENTRIES IN USE catches up to the NUMBER LIST ENTRIES, new modify lock requests will not be granted, and transactions will begin to fail with -904s.

LOCK TABLE

A 'POOL' OF HASH POINTERS
CONTAINS ENTRIES FOR ALL READ AND MODIFY LOCKS
USED FOR GLOBAL LOCK CONTENTION
It is very difficult to keep track of how full a randomly populated table is, so the only way to track its utilization is indirectly. In DB2 data sharing, that tracking mechanism is false contention
SIZE OF EACH INDIVIDUAL ENTRY SET BY MAXUSRS PARAMETER IN IRLMPROC OF FIRST IRLM JOIN THE GROUP
NUMBER OF MEMBER IN THE GROUP

XCF

其实主要的目的就是提供INTER-SUBSYTEM之间通讯的API，主要的功能有三个：

1. Signaling Services provide a method for communicating between members of the same XCF roup, which is a set of related members defined to XCF by a multisystem application. This function supports the single system image concept of a parallel sysplex, both by providing communication etween subsystems, such as DB2 members, as well as z/OS components in each image.
2. Group Services allow multisystem applications to define groups and their members to XCF. Group services also provide the means for requesting information about the other members in the same XCF group, using SETXCF commands.
3. Status Monitoring Services of XCF provide the means for monitoring the status of other z/OS systems in the same sysplex.

   XES

   提供了不同的SUBSYSTEM访问CF的API。XES本身仅支持两种锁类型，即S|X类型。但是DB2 SUBSYTEMB本身支持多种不同类型的锁，这就带来一个问题，这些锁类型在上送到CF时，应该如何映射，DB2 V7 Protocol Level 1和V8 Protocol Level 2的处理差别比较大，当然这种变化的目的还是在保证能完全检测到锁冲突的前提下，减少上送到CF锁的请求次数，优化全局锁检测机制。V7到V8优化的出发点还是进一步的解决XES锁映射问题,V8到V10甚至V11基本上无变化，可见V8优化的意义之重大。
   DB2在DATA SHARING 下是显式的三级锁结构，TWO LEVELS OF APPLICATION LOCKS外加DB2 member级别的PAGE SET P LOCK.
   DB2 PAGE SET P-LOCK的状态变化如下图：

   action	DB2 member interest page set p lock	STAE CHANGE
   open	PP-LOCK S/IS	NONE->R/O
   PSEUDO OPEN	PP-LOCK IX/SIX/X/U	R/O->R/W
   PSEDUO CLOSE	PP-LOCK S/IS	R/W->R/O
   Physical CLOSE	NONE	R/O->NONE

XES CONTENTION

产生的根源为XES本身设计时仅支持两种锁类型，
?XES Contention = XES-level resource contention as XES
only understands S or X
eg member 1 asking for IX and member 2 for IS
Big relief in V8

FALSE CONTENTION

产生的根源主要是由于DB2 lock table定义的太小或是hash算法离散型不够好，导致不同的DB2资源hash到同一lock table hash class上而产生的冲突，可以通过放大CF lock table来解决，因此FALSE CONTENTION标志着 CF LOCK STRUCTURE的大小是否合适。

GLOBAL CONTENTION

GLOBAL CONTENTION 本身包含XES CONTENTION,FALSE CONTENTION,REAL CONTENTION,其中REAL CONTENTION 确实为资源冲突，如资源热点，其它的两种冲突是由于引入DATA SHARING 的OVERHEAD

根据上面的介绍，下面给出A DB2 Performance Tuning Roadmap-2data sharing BLOCK的一个展示

L-LOCK(SHARING)部分表明了数据共享的