1
1/11/05 1
3. C oncurrency C ontrol
for Transactions
Part One
Transaction Processing
1
/
11
/
05 2
Outline
1. A Simple System M odel
2. Serializab ility Theory
3. Synchronization R equirem ents
for R eco verability
4. Tw o-Ph ase Locking
5. Preserving Transaction Handshakes
6. Implem enting Two-P hase Locking
7. D eadlocks
1
/
11
/
05 3
3.1 A Simple System M odel
•Goal - Ensure serializable (SR) executions
•Implem entation technique -D elay operations
that w ould lead to non-SR results (e.g. se t locks
on sh ared data)
•For good perform ance minimize overhead and
delay from syn ch ronization operations
•First, we’ll study how to get correct (SR) results
•Then, we’ll study perform ance implications
(mostly in Part Tw o)
1
/
11
/
05 4
A ssu m p tion -A tom ic Operations
•We will synchronize Reads and Writes.
•We must therefore assu m e they’re atom ic
–else we’d have to synchronize the finer-grained
operations that implem en t Read and Write
• R ead (x) -returns the current value of x in the DB
•Write(x, val) overwrites allof x (the w hole page)
•This assu m p tion of atom ic operations is what
allow s us to ab stract executions as sequences of
reads and writes (without lo ss of information).
–Otherwise, what would wk[x] r
i
[x] mean?
•Also, co m m it (c
i
) and abort (a
i
) are atom ic
1
/
11
/
05 5
System M odel
Tran saction 1 Tran saction N
Start, Commit, Abort
Read(x), Write(x)
Data
Manager
Database
Tran saction 2
1
/
11
/
05 6
3.2 Serializability Theory
•The theory is based on m odeling executions as
histories, su ch as
H1= r
1[x] r
2[x] w 1[x] c
1w2[y] c
2
•First, characterize a concurrency control
algorithm by the properties of histories it allow s.
•Then prove that any history having these
properties is SR
•Why bother? It helps you understand why
concurrency control algorithm s work.
Docsity.com
