Real-Time Database: Scheduling and Concurrency Control

1
Real-Time Database Scheduling and Concurrency
Control

2
Paper Sources (1)

• On Real-Time Databases Concurrency Control and
  Scheduling
• Survey paper
• Traditional Database Concurrency Control
• RTDB Scheduling and Concurrency Control
• New semantics for RTDB

3
Paper Sources (2)

• Scheduling Real-Time Transactions A Performance
  Evaluation
• Performance comparison paper
• Admission Control
• Scheduling
• Concurrency Control
• Main memory based vs. Disk based Database

4
Database and Real-Time (1)

• Traditional Database
• Data Consistency
• Transaction Serializability
• Traditional Concurrency Control mechanisms
• Locking based
• Timestamp Ordering
• Serialization Graph

5
Database and Real-Time (2)

• Real-Time System
• Hard real-time
• Soft real-time
• Scheduling Algorithms
• Earliest Deadline First (ED)
• Least Slack Time (LS)
• First Come First Serve (FCFS)
• Others

6
Preliminary about RTDB (1)

• Difference between traditional database and RTDB
• Constraints on data items
• Constraints on transactions
• Different performance goals
• Different correctness definition
• Difference between other real-time system and
  RTDB
• Data consistency
• Predictability of data requirement

7
Preliminary about RTDB (2)

• Problems with RT-database
• hard to predict execution time
• Hard to predict data access requirement
• I/O
• Integration with Concurrency Control

8
Preliminary about RTDB (3)

• Hard Real-time System
• Special requirements
• Guarantee to meet deadlines
• Possible solutions
• Advanced knowledge of resource and data
  requirement
• Relaxed serializability requirement
• Relaxed data consistency requirement

9
Data Consistency (1)

• External Consistency
• Relationship between data and the real world
• Temporal Consistency
• Relationship between data

10
Data Consistency (2)

• Periodic updates
• Convert consistency metrics into upper lower
  bounds for update transaction periods
• Trading serializability for external consistency

11
Data Consistency (3)

• Transaction Compatibility Table (TCT)
• Better for pre-run-time analysis
• Epsilon Serializability (ESR)
• Limited degree of inconsistency

12
Concurrency Control (1)

• Conflict Detection
• Conflict Resolution
• Serialization Rule and Order
• Run Policy

13
Concurrency Control (2)

• Conflict Detection
• Lock-based scheme
• This is the CC scheme most RTDBs use
• Read lock, Write lock and Weak lock
• Granularity of locks
• Timestamp Ordering
• Serialization Graph

14
Concurrency Control (3)

• Conflict Resolution
• Choices
• blocking, abort, multiversioning,, dynamically
  adjusting the serialization order
• 2PL and Variations
• Wait or Abort?
• With or without Priority?
• Cascading blocking
• Multiversioning
• Two versions
• Multiversions

15
Concurrency Control (4)

• Conflict Resolution
• OCC and variations
• Resource waste
• Pure OCC delayed abort
• Broadcast OCC immediate abort
• Combination of OCC and 2PL
• Locking with Deferred Blocking (LDB)
• Nonblocking phase and blocking phase

16
Concurrency Control (5)

• Serialization Rule and Order
• With or without readjusting serialization order?
• Start time order TSC
• Completion time order OCC
• Access order 2PL
• Dynamic order Dynamic timestamp allocation,
  Timestamp interval allocation
• A mixed read-write behavior
• Time-stamp History

17
Concurrency Control (6)

• Run Policy
• Buffer Retention effect
• Different Concurrency Control schemes could be
  applied to the first-run and rerun transactions

18
Scheduling RT Transactions (1)

• Old methods fit new scenario?
• Locking mechanisms
• Timestamp Ordering
• OCC
• Conflict!
• Single conflict
• Conflict with a set
• Fairness!
• Long transactions
• Almost finished transactions

19
Scheduling RT Transactions (2)

• How to solve the problem?
• Priority
• Priority from the scheduling point of view
• A mapping problem
• EDF
• Weighted Priority

20
Scheduling RT Transactions (3)

• Locking Mechanisms
• Priority Inversion
• Priority Inheritance
• Priority Ceiling
• Convex Ceiling
• 2PL with High Priority (2PL-HP)
• Conditional Priority Inheritance (CPI)
• Other Mechanisms
• OCC, Multiversioning, etc.

21
Scheduling RT Transactions (4)

• Interaction between system load and scheduling
• Lower load situation
• EDF
• High load situation
• LS
• Adaptive Earliest Deadline (AED)
• Adaptive Earliest Virtual Deadline (AEVD)

22
Scheduling RT Transactions (5)

• I/O Issues in RTDB
• Disk Access Scheduling
• Elevator Scan
• EDF
• Combination of two mechanisms
• Read/Write policy
• Non-deferred Write
• Deferred Write

23
Scheduling RT Transactions (6)

• Run Policy
• Prefetching
• First Run
• Get all the data into main memory
• Estimation
• Second Run
• Interaction with other re-run transactions
• Interaction with first-run transactions

24
Performance Evaluation (1)

• Evaluation Metrics
• Overload management
• AE/ NT/ FD
• Priority Assignment
• FCFS/ ED/ LS
• Concurrency Control
• W/ WP/ HP/ CR
• I/O Scheduling
• FIFO/ P

25
Performance Evaluation (2)

• System model
• Single processor
• Disk-based database
• Main memory buffer
• High Workload
• Transaction model
• Release time r
• Deadline d
• Runtime estimate e

26
Performance Evaluation (3)

• Database Model
• Page-based granularity
• Promise Serializability
• Locking mechanism
• No write-back to disk when abort

27
Performance Evaluation (4)

• Overload Management
• Observant vs. Predictive
• All Eligible (AE)
• Not Tardy (NT)
• Feasible Deadlines (FD)
• I/O Scheduling
• First In First Out (FIFO)
• Priority

28
Performance Evaluation (5)

• Priority Assignment
• deadline vs. service
• First Come First Serve (FCFS)
• Earliest Deadline (ED)
• Least Slack (LS)
• Static evaluation
• Continuous Evaluation

29
Performance Evaluation (6)

• Concurrency Control
• Priority Inversion
• Wait or Abort
• Wait (W)
• Wait Promote (WP)
• High Priority (HP)
• Conditional Restart (CR)
• One-to-one conflict assumption

30
Performance Evaluation (7)

• Simulation Model

31
Result for Memory-Resident DB (1)

• For Priority Assignments
• ED performs better at lower load
• LS performs better at higher load
• LS permits a greater degree of concurrency than
  ED
• However, higher concurrency might lead to more
  lock conflicts

32
Result for Memory-Resident DB (2)

• Run Time Estimation Error
• Random
• Biased
• For overload management FD
• For priority policy LS
• Static evaluation
• Dynamic evaluation
• For concurrency control CR
• E0 (WP)
• EgtgtR (HP)

33
Result for Memory-Resident DB (3)

• Other points
• Importance of overload management
• WP and CR matches LS well
• HP causes too many aborts at high load
• LS/W and LS/HP are more sensitive to rate of
  conflict
• Cost of restart

34
Result for Disk-Resident DB

• Consistent with MM experiment results
• Priority inheritance preferred
• For Sudden Workload Increase, ED/HP is the best
• Effect of memory size