Question? Leave a message!




Concurrency Control

Concurrency Control
Chapter 9: Concurrency Control • Concurrency, Conflicts, and Schedules • Locking Based Algorithms • Timestamp Ordering Algorithms • Deadlock Management Acknowledgements: I am indebted to Arturas Mazeika for providing me his slides of this course. DDB 2008/09 J. Gamper Page 1Concurrency • Concurrency control is the problem of synchronizing concurrent transactions (i.e., order the operations of concurrent transactions) such that the following two properties are achieved: – the consistency of the DB is maintained – the maximum degree of concurrency of operations is achieved • Obviously, the serial execution of a set of transaction achieves consistency, if each single transaction is consistent DDB 2008/09 J. Gamper Page 2Conflicts • Conflicting operations: Two operationsO (x) andO (x) of transactionsT andT ij kl i k are in conflict iff at least one of the operations is a write, i.e., – O =read(x) andO =write(x) ij kl – O =write(x) andO =read(x) ij kl – O =write(x) andO =write(x) ij kl • Intuitively, a conflict between two operations indicates that their order of execution is important. • Read operations do not conflict with each other, hence the ordering of read operations does not matter. • Example: Consider the following two transactions T : Read(x) T : Read(x) 1 2 x←x+1 x←x+1 Write(x) Write(x) Commit Commit – To preserve DB consistency, it is important that theread(x) of one transaction is not betweenread(x) andwrite(x) of the other transaction. DDB 2008/09 J. Gamper Page 3Schedules • A schedule (history) specifies a possibly interleaved order of execution of the operations O of a set of transactionsT =T ,T ,...,T , whereT is specified by a partial 1 2 n i order(Σ ,≺ ). A schedule can be specified as a partial order overO, where i i S n – Σ = Σ T i i=1 S n –≺ ⊇ ≺ T i i=1 – For any two conflicting operationsO ,O ∈Σ , eitherO ≺ O or ij kl T ij T kl O ≺ O kl T ij DDB 2008/09 J. Gamper Page 4Schedules . . . • Example: Consider the following two transactions T : Read(x) T : Read(x) 1 2 x←x+1 x←x+1 Write(x) Write(x) Commit Commit – A possible schedule overT =T ,T can be written as the partial order 1 2 S =Σ ,≺ , where T T Σ =R (x),W (x),C ,R (x),W (x),C T 1 1 1 2 2 2 ≺ =(R ,W ),(R ,C ),(W ,C ), T 1 1 1 1 1 1 (R ,W ),(R ,C ),(W ,C ), 2 2 2 2 2 2 (R ,W ),(W ,W ),... 2 1 1 2 DDB 2008/09 J. Gamper Page 5Schedules . . . • A schedule is serial if all transactions inT are executed serially. • Example: Consider the following two transactions T : Read(x) T : Read(x) 1 2 x←x+1 x←x+1 Write(x) Write(x) Commit Commit – The two serial schedules areS =Σ ,≺ andS =Σ ,≺, where 1 1 1 2 2 2 Σ =Σ =R (x),W (x),C ,R (x),W (x),C 1 2 1 1 1 2 2 2 ≺ =(R ,W ),(R ,C ),(W ,C ),(R ,W ),(R ,C ),(W ,C ), 1 1 1 1 1 1 1 2 2 2 2 2 2 (C ,R ),... 1 2 ≺ =(R ,W ),(R ,C ),(W ,C ),(R ,W ),(R ,C ),(W ,C ), 2 1 1 1 1 1 1 2 2 2 2 2 2 (C ,R ),... 2 1 • We will also use the following notation: –T ,T=R (x),W (x),C ,R (x),W (x),C 1 2 1 1 1 2 2 2 –T ,T=R (x),W (x),C ,R (x),W (x),C 2 1 2 2 2 1 1 1 DDB 2008/09 J. Gamper Page 6Serializability • Two schedules are said to be equivalent if they have the same effect on the DB. • Conflict equivalence: Two schedulesS andS defined over the same set of 1 2 transactionsT =T ,T ,...,T are said to be conflict equivalent if for each pair 1 2 n of conflicting operationsO andO , wheneverO O thenO O . ij ij 1 ij 2 kl kl kl – i.e., conflicting operations must be executed in the same order in both transactions. • A concurrent schedule is said to be (conflict)serializable iff it is conflict equivalent to a serial schedule • A conflictserializable schedule can be transformed into a serial schedule by swapping nonconflicting operations • Example: Consider the following two schedules T : Read(x) 1 T : Read(x) 2 x←x+1 x←x+1 Write(x) Write(x) Write(z) Commit Commit – The scheduleR (x),W (x),R (x),W (x),W (z),C ,C is 1 1 2 2 1 2 1 conflictequivalent toT ,T but not toT ,T 1 2 2 1 DDB 2008/09 J. Gamper Page 7Serializability . . . • The primary function of a concurrency controller is to generate a serializable schedule for the execution of pending transactions. • In a DDBMS two schedules must be considered – Local schedule – Global schedule (i.e., the union of the local schedules) • Serializability in DDBMS – Extends in a straightforward manner to a DDBMS if data is not replicated – Requires more care if data is replicated: It is possible that the local schedules are serializable, but the mutual consistency of the DB is not guaranteed. ∗ Mutual consistency: All the values of all replicated data items are identical • Therefore, a serializable global schedule must meet the following conditions: – Local schedules are serializable – Two conflicting operations should be in the same relative order in all of the local schedules they appear ∗ Transaction needs to be run on each site with the replicated data item DDB 2008/09 J. Gamper Page 8Serializability . . . • Example: Consider two sites and a data itemx which is replicated at both sites. T : Read(x) T : Read(x) 1 2 x←x+5 x←x∗10 Write(x) Write(x) – Both transactions need to run on both sites – The following two schedules might have been produced at both sites (the order is implicitly given): ∗ Site1: S =R (x),W (x),R (x),W (x) 1 1 1 2 2 ∗ Site2: S =R (x),W (x),R (x),W (x) 2 2 2 1 1 – Both schedules are (trivially) serializable, thus are correct in the local context – But they produce different results, thus violate the mutual consistency DDB 2008/09 J. Gamper Page 9Concurrency Control Algorithms • Taxonomy of concurrency control algorithms – Pessimistic methods assume that many transactions will conflict, thus the concurrent execution of transactions is synchronized early in their execution life cycle ∗ TwoPhase Locking (2PL) · Centralized (primary site) 2PL · Primary copy 2PL · Distributed 2PL ∗ Timestamp Ordering (TO) · Basic TO · Multiversion TO · Conservative TO ∗ Hybrid algorithms – Optimistic methods assume that not too many transactions will conflict, thus delay the synchronization of transactions until their termination ∗ Lockingbased ∗ Timestamp orderingbased DDB 2008/09 J. Gamper Page 10Locking Based Algorithms • Lockingbased concurrency algorithms ensure that data items shared by conflicting operations are accessed in a mutually exclusive way. This is accomplished by associating a “lock” with each such data item. • Two types of locks (lock modes) – read lock (rl) – also called shared lock – write lock (wl) – also called exclusive lock • Compatibility matrix of locks rl (x) wl (x) i i rl (x) compatible not compatible j wl (x) not compatible not compatible j • General locking algorithm 1. Before using a data itemx, transaction requests lock forx from the lock manager 2. Ifx is already locked and the existing lock is incompatible with the requested lock, the transaction is delayed 3. Otherwise, the lock is granted DDB 2008/09 J. Gamper Page 11Locking Based Algorithms • Example: Consider the following two transactions T : Read(x) T : Read(x) 1 2 x←x+1 x←x∗2 Write(x) Write(x) Read(y) Read(y) y←y−1 y←y∗2 Write(y) Write(y) – The following schedule is a valid lockingbased schedule (lr (x) indicates the i release of a lock onx): S =wl (x),R (x),W (x),lr (x) 1 1 1 1 wl (x),R (x),W (x),lr (x) 2 2 2 2 wl (y),R (y),W (y),lr (y) 2 2 2 2 wl (y),R (y),W (y),lr (y) 1 1 1 1 – However,S is not serializable ∗ S cannot be transformed into a serial schedule by using only nonconflicting swaps ∗ The result is different from the result of any serial execution DDB 2008/09 J. Gamper Page 12TwoPhase Locking (2PL) • Twophase locking protocol – Each transaction is executed in two phases ∗ Growing phase: the transaction obtains locks ∗ Shrinking phase: the transaction releases locks – The lock point is the moment when transitioning from the growing phase to the shrinking phase DDB 2008/09 J. Gamper Page 13TwoPhase Locking (2PL) . . . • Properties of the 2PL protocol – Generates conflictserializable schedules – But schedules may cause cascading aborts ∗ If a transaction aborts after it releases a lock, it may cause other transactions that have accessed the unlocked data item to abort as well • Strict 2PL locking protocol – Holds the locks till the end of the transaction – Cascading aborts are avoided DDB 2008/09 J. Gamper Page 14TwoPhase Locking (2PL) . . . • Example: The scheduleS of the previous example is not valid in the 2PL protocol: S =wl (x),R (x),W (x),lr (x) 1 1 1 1 wl (x),R (x),W (x),lr (x) 2 2 2 2 wl (y),R (y),W (y),lr (y) 2 2 2 2 wl (y),R (y),W (y),lr (y) 1 1 1 1 – e.g., afterlr (x) (in line 1) transactionT cannot request the lockwl (y) (in line 4). 1 1 1 – Valid schedule in the 2PL protocol S =wl (x),R (x),W (x), 1 1 1 wl (y),R (y),W (y),lr (x),lr (y) 1 1 1 1 1 wl (x),R (x),W (x), 2 2 2 wl (y),R (y),W (y),lr (x),lr (y) 2 2 2 2 2 DDB 2008/09 J. Gamper Page 152PL for DDBMS • Various extensions of the 2PL to DDBMS • Centralized 2PL – A single site is responsible for the lock management, i.e., one lock manager for the whole DDBMS – Lock requests are issued to the lock manager – Coordinating transaction manager (TM at site where the transaction is initiated) can make all locking requests on behalf of local transaction managers • Advantage: Easy to implement • Disadvantages: Bottlenecks and lower reliability • Replica control protocol is addi tionally needed if data are repli cated (see also primary copy 2PL) DDB 2008/09 J. Gamper Page 162PL for DDBMS . . . • Primary copy 2PL – Several lock managers are distributed to a number of sites – Each lock manager is responsible for managing the locks for a set of data items – For replicated data items, one copy is chosen as primary copy, others are slave copies – Only the primary copy of a data item that is updated needs to be writelocked – Once primary copy has been updated, the change is propagated to the slaves • Advantages – Lower communication costs and better performance than the centralized 2PL • Disadvantages – Deadlock handling is more complex DDB 2008/09 J. Gamper Page 172PL for DDBMS . . . • Distributed 2PL – Lock managers are distributed to all sites – Each lock manager responsible for locks for data at that site – If data is not replicated, it is equivalent to primary copy 2PL – If data is replicated, the ReadOneWriteAll (ROWA) replica control protocol is implemented ∗ Read(x): Any copy of a replicated itemx can be read by obtaining a read lock on the copy ∗ Write(x): All copies ofx must be writelocked beforex can be updated • Disadvantages – Deadlock handling more complex – Communication costs higher than primary copy 2PL DDB 2008/09 J. Gamper Page 182PL for DDBMS . . . • Communication structure of the distributed 2PL – The coordinating TM sends the lock request to the lock managers of all participating sites – The LMs pass the operations to the data processors – The end of the operation is signaled to the coordinating TM DDB 2008/09 J. Gamper Page 19Timestamp Ordering • Timestampordering based algorithms do not maintain serializability by mutual exclusion, but select (a priori) a serialization order and execute transactions accordingly. – TransactionT is assigned a globally unique timestampts(T ) i i – Conflicting operationsO andO are resolved by timestamp order, i.e.,O is ij ij kl executed beforeO iffts(T )ts(T ). i kl k • To allow for the scheduler to check whether operations arrive in correct order, each data item is assigned a write timestamp (wts) and a read timestamp (rts): – rts(x): largest timestamp of any read onx – wts(x): largest timestamp of any write onx • Then the scheduler has to perform the following checks: – Read operation,R (x): i ∗ Ifts(T )wts(x): T attempts to read overwritten data; abortT i i i ∗ Ifts(T )≥wts(x): the operation is allowed andrts(x) is updated i – Write operations,W (x): i ∗ Ifts(T )rts(x): x was needed before by other transaction; abortT i i ∗ Ifts(T )wts(x): T writes an obsolete value; abortT i i i ∗ Otherwise, executeW (x) i DDB 2008/09 J. Gamper Page 20Timestamp Ordering . . . • Generation of timestamps (TS) in a distributed environment – TS needs to be locally and globally unique and monotonically increasing – System clock, incremental event counter at each site, or global counter are unsuitable (difficult to maintain) – Concatenate local timestamp/counter with a unique site identifier: local timestamp, site identifier ∗ site identifier is in the least significant position in order to distinguish only if the local timestamps are identical • Schedules generated by the basic TO protocol have the following properties: – Serializable – Since transactions never wait (but are rejected), the schedules are deadlockfree – The price to pay for deadlockfree schedules is the potential restart of a transaction several times DDB 2008/09 J. Gamper Page 21Timestamp Ordering . . . • Basic timestamp ordering is “aggressive”: It tries to execute an operation as soon as it receives it • Conservative timestamp ordering delays each operation until there is an assurance that it will not be restarted, i.e., that no other transaction with a smaller timestamp can arrive – For this, the operations of each transaction are buffered until an ordering can be established so that rejections are not possible • If this condition can be guaranteed, the scheduler will never reject an operation • However, this delay introduces the possibility for deadlocks DDB 2008/09 J. Gamper Page 22Timestamp Ordering . . . • Multiversion timestamp ordering – Write operations do not modify the DB; instead, a new version of the data item is created: x ,x ,...,x 1 2 n – R (x) is always successful and is performed on the appropriate version ofx, i.e., the i version ofx (sayx ) such thatwts(x ) is the largest timestamp less thants(T ) v v i – W (x) produces a new versionx withts(x )=ts(T ) if the scheduler has not i w w i yet processed anyR (x ) on a versionx such that j r r ts(T )rts(x ) i r i.e., the write is too late. – Otherwise, the write is rejected. DDB 2008/09 J. Gamper Page 23Timestamp Ordering . . . • The previous concurrency control algorithms are pessimistic • Optimistic concurrency control algorithms – Delay the validation phase until just before the write phase – T run independently at each site on local copies of the DB (without updating the DB) i – Validation test then checks whether the updates would maintain the DB consistent: ∗ If yes, all updates are performed ∗ If one fails, allT ’s are rejected i • Potentially allow for a higher level of concurrency DDB 2008/09 J. Gamper Page 24Deadlock Management • Deadlock: A set of transactions is in a deadlock situation if several transactions wait for each other. A deadlock requires an outside intervention to take place. • Any lockingbased concurrency control algorithm may result in a deadlock, since there is mutual exclusive access to data items and transactions may wait for a lock • Some TObased algorihtms that require the waiting of transactions may also cause deadlocks • A Waitfor Graph (WFG) is a useful tool to identify deadlocks – The nodes represent transactions – An edge fromT toT indicates thatT is waiting forT i j i j – If the WFG has a cycle, we have a deadlock situation DDB 2008/09 J. Gamper Page 25Deadlock Management . . . • Deadlock management in a DDBMS is more complicate, since lock management is not centralized • We might have global deadlock, which involves transactions running at different sites • A Local WaitforGraph (LWFG) may not show the existence of global deadlocks • A Global Waitfor Graph (GWFG), which is the union of all LWFGs, is needed DDB 2008/09 J. Gamper Page 26Deadlock Management . . . • Example: AssumeT andT run at site 1,T andT run at site 2, and the following 1 2 3 4 waitfor relationships between them: T →T →T →T →T . This deadlock 1 2 3 4 1 cannot be detected by the LWFGs, but by the GWFG which shows intersite waiting. – Local WFG: – Global WFG: DDB 2008/09 J. Gamper Page 27Deadlock Prevention • Deadlock prevention: Guarantee that deadlocks never occur – Check transaction when it is initiated, and start it only if all required resources are available. – All resources which may be needed by a transaction must be predeclared • Advantages – No transaction rollback or restart is involved – Requires no runtime support • Disadvantages – Reduced concurrency due to preallocation – Evaluating whether an allocation is safe leads to added overhead – Difficult to determine in advance the required resources DDB 2008/09 J. Gamper Page 28Deadlock Avoidance • Deadlock avoidance: Detect potential deadlocks in advance and take actions to ensure that a deadlock will not occur. Transactions are allowed to proceed unless a requested resource is unavailable • Two different approaches: – Ordering of data items: Order data items and sites; locks can only be requested in that order (e.g., graphbased protocols) – Prioritize transactions: Resolve deadlocks by aborting transactions with higher or lower priority. The following schemes assume thatT requests a lock hold byT : i j ∗ WaitDie Scheme: ifts(T )ts(T ) thenT waits elseT dies i j i i ∗ WoundWait Scheme: ifts(T )ts(T ) thenT wounds (aborts) elseT waits i j j i • Advantages – More attractive than prevention in a database environment – Transactions are not required to request resources a priori • Disadvantages – Requires run time support DDB 2008/09 J. Gamper Page 29Deadlock Detection • Deadlock detection and resolution: Transactions are allowed to wait freely, and hence to form deadlocks. Check global waitfor graph for cycles. If a deadlock is found, it is resolved by aborting one of the involved transactions (also called the victim). • Advantages – Allows maximal concurrency – The most popular and beststudied method • Disadvantages – Considerable amount of work might be undone • Topologies for deadlock detection algorithms – Centralized – Distributed – Hierarchical DDB 2008/09 J. Gamper Page 30Deadlock Detection . . . • Centralized deadlock detection – One site is designated as the deadlock detector (DDC) for the system – Each scheduler periodically sends its LWFG to the central site – The site merges the LWFG to a GWFG and determines cycles – If one or more cycles exist, DDC breaks each cycle by selecting transactions to be rolled back and restarted • This is a reasonable choice if the concurrency control algorithm is also centralized DDB 2008/09 J. Gamper Page 31Deadlock Detection . . . • Hierarchical deadlock detection – Sites are organized into a hierarchy – Each site sends its LWFG to the site above it in the hierarchy for the detection of deadlocks – Reduces dependence on centralized detection site DDB 2008/09 J. Gamper Page 32Deadlock Detection . . . • Distributed deadlock detection – Sites cooperate in deadlock detection – The local WFGs are formed at each site and passed on to the other sites. – Each local WFG is modified as follows: ∗ Since each site receives the potential deadlock cycles from other sites, these edges are added to the local WFGs ∗ i.e., the waiting edges of the local WFG are joined with waiting edges of the external WFGs – Each local deadlock detector looks for two things: ∗ If there is a cycle that does not involve the external edge, there is a local deadlock which can be handled locally ∗ If there is a cycle involving external edges, it indicates a (potential) global deadlock. DDB 2008/09 J. Gamper Page 33Conclusion • Concurrency orders the operations of transactions such that two properties are achieved: (i) the database is always in a consistent state and (ii) the maximum concurrency of operations is achieved • A schedule is some order of the operations of the given transactions. If a set of transactions is executed one after the other, we have a serial schedule. • There are two main groups of serializable concurrency control algorithms: locking based and timestamp based • A transaction is deadlocked if two or more transactions are waiting for each other. A Waitfor graph (WFG) is used to identify deadlocks • Centralized, distributed, and hierarchical schemas can be used to identify deadlocks DDB 2008/09 J. Gamper Page 34