Logical and Physical Versioning in Main Memory Databases

1
Logical and Physical Versioning in Main Memory
Databases

• -Dali Project Team-
• from the proceedings of the 23rd VLDB conference .

2
Introduction(I)

• MMDB improves performance
• dispensing with the buffer manager
• tuning algorithms to the flat storage hierarchy
• reduced cost of indirection
• But need to minimize space usage
• MMDB is suitable for real-time applications
• read-only transactions response time
• multi-version concurrency control

3
Introduction(II)

• By avoiding latches, significant performance
  gains can be obtained for read-only transactions
• Logical versioning
• locks are eliminated by versioning of data items
• reduces the storage space overhead of keeping
  track of versions
• Physical versioning
• applied to the access paths to data items
• latches are eliminated

4
Logical Versioning(I)

• Update transactions
• follows two-phase locking protocol
• at the commit time, assigned new timestamp
• Read-only transaction
• get timestamp by reading the global logical
  timestamp
• A version can be aged safely if ...

5
Logical Versioning(II)

• Assigning Timestamps
• Version List Entries
• space for versions is dynamically allocated
• data and VLE are stored on distinct segments

6
Logical Versioning(III)

• Interaction with Indices

7
Physical Versioning

• A technique that permits read-only transactions
  to access data structures without getting any
  latches or locks,even while other update
  transactions are updating the data structure
• performed as follows
• copy component N to N
• update on the N
• atomically update the pointer to N to point to N
• physical aging

8
T-Tree

• Overview of T-tree LC86
• latch and versioning
• Find / Insert / Balancing

9
Performance Results(I)

• The effects of physical versioning and latching
  overheads on performance
• tree latch with no versioning
• node latch with no versioning
• tree latch with physical versioning
• node latch with physical versioning
• T-tree has 10 keys in each node, key range is 0
  2,000,000
• insert percentage 1 75, running time 5min.

10
Performance Results(II)
11
Conclusion

• Presented a design for multi-version concurrency
  control and index management in a MMDB
• showed how to support real time performance for
  read-only transactions
• applied it to the T-tree MMDB index structure
  demonstrated performance improvements