Concurrency control protocols in MDAS An evaluation

1

• Concurrency control protocols in MDAS ----An
  evaluation
• By Alan (Xing Gao)
• Advised by Prof. Ali R. Hurson

2
Table of content

• I MDAS structure.
• II Concurrency protocols.
• III Simulation result and analysis.
• IV Effect of resubmission.
• V Future work.

3
MDAS structure

• A centralized Multi-database system.

Global Transactions
Global MDBS Manager
Global Sub-transaction

Local DBMS 1
Local DBMS 2
Local DBMS 3
Local DBMS m
Local Transactions
Local Transactions
Local Transactions
Local Transactions
4
MDAS structure

• Problem
• Lack of scalability.
• Global transaction manager becomes a bottleneck
  in communication and locking table access.
• Solution
• Summary Schemas Model (SSM)
• Distribute the manager functions into a
  hierarchy of SSM nodes and therefore avoid the
  bottleneck.

5
MDAS structure SSM

• Summary Schemas Model (SSM)
• Uses a hierarchical structure that provides an
  incrementally concise view of data in the form of
  summary schemas.

6
MDAS structure SSM Hierarchy

• SSM hierarchy used in my simulator.
• 5 levels hierarchy
• 10 database nodes
• 11 SSM nodes

7
MDAS structure SSM Hierarchy
8
MDAS structure SSM Hierarchy

• Semantic summarization
• Db0 employee salary
• ssm6 salary
• Db1 employer salary
• ssm3 Factory database
• Db2 wholesale data
• ssm7 sale
• Db3 retail data

9
MDAS structure An Example

• GT0 needs data from db0 and db3.
• GT1 needs data from db5 and db7.
• GT2 needs data from db8 and db9.
• GT3 needs data from db1 and db5.

10
MDAS structure Execution
11
Concurrency Protocols

• Four protocols are simulated
• Forced Conflict
• Site Graph
• Potential Conflict Graph
• V-lock

12
Protocols Forced Conflict

• Forced Conflict
• Add a special data item called ticket at each
  local site.
• Requires each global sub-transaction to access
  the ticket at its site.
• The tickets obtained by the sub-transactions of
  multi-database transactions determine their
  relative serialization order.
• Issue next operation only if the previous one
  succeeds.

13
Protocols Site Graph

• Site Graph
• A site graph on a transaction is a connection
  among all sites where of of its sub-transactions
  is executing.
• Each SSM node maintains a site graph for all
  global transactions under its control.
• For each operation issued to a local site by an
  SSM, an edge is added to the site graph of this
  transaction.
• If SSM cannot find a site to execute the
  operation without leading to a cycle in the site
  graph, it aborts the transaction.
• The acyclicness of the site graph will guarantee
  the correctness (serializability) of global
  execution.
• Issue next operation only if the previous one
  succeeds.

14
Protocols Potential Conflict Graph

• Potential Conflict Graph
• Assume that the local DBMSs use
  two-phase-locking.
• Each SSM maintains a global-wait-for-graph, a
  commit-graph , a wait-for-commit-graph, and a
  potential-conflict-graph.
• The acyclicness of the four graphs will guarantee
  the correctness (serializability) of global
  execution.
• Issue next operation only if the previous one
  succeeds.

15
Protocols V-Lock

• V-Lock
• Assume that the local DBMSs use
  two-phase-locking.
• Each SSM maintains a global-wait-for-graph, a
  commit-graph , wait-for-commit-graph, and a
  potential-conflict-graph.
• The acyclicness of the four graphs will guarantee
  the correctness (serializability) of global
  execution.
• Dynamically adjust acknowledgement frequency
  between coordinator and local sites.
• Can issue the whole sub-transaction at once.

16
Simulation Simulator

• Simulator
• The programs are written in C language.
• Using CSIM simulation package.
• Assume busy global and local systems.
• Assume a reliable connection.
• No disconnection, fixed delay.
• All four concurrency protocol share the same
  parameters in global system and local systems.

17
Simulation Local Parameters

• Default local parameters

18
Simulation Global Parameters
19
Simulation Result

• Performance metrics
• Completion rate
• Global throughput
• Average response time
• Local CPU utilization
• Local IO utilization
• Communication utilization

20
Simulation Completion Rate
21
Simulation Completion Rate

• V-Lock gt PCGraph gt Forced Conflict gt SG
• when global concurrency degree lt50
• All four protocols have poor completion rates
  (lt20) when global concurrency degree gt50

22
Simulation Global Throughput
23
Simulation Global Throughput

• V-lock has the highest throughput when the
  concurrency degree is below 7.
• Site Graph has the highest throughput when the
  concurrency degree is above 7.
• (at expense of low completion rate. lt40)
• Forced Conflict and Potential Conflict Graph have
  relatively steady throughputs.

24
Simulation Avg. Response Time
25
Simulation Avg. Response Time

• Response time for Site Graph does not change much
  as global concurrency degree increases.
• All other three protocols experience an increase
  and then a drop in their response time.
• Response time for V-lock is very sensitive to the
  global concurrency degree.
• Majority aborts in Site Graph is because of
  possible deadlock.
• Many aborts in V-lock are because of timeout.

26
Simulation Local I/O Utilization
27
Simulation Local I/O Utilization

• No direct relation to global throughput.
• Low I/O utilization indicates more global
  sub-transactions active/waiting at local sites.
• Site Graph have a high I/O utilization rate than
  others
• Less global sub-transactions at local sites.

28
Simulation Local CPU Utilization
29
Simulation Local CPU Utilization

• CPU utilization and I/O utilization have similar
  curves.
• No direct relation to global throughput.
• Low CPU utilization indicates more global
  sub-transactions active/waiting at local sites.
• Site Graph have a high CPU utilization rate than
  others
• Less global sub-transactions at local sites.

30
Simulation Local Comm Utilization
31
Simulation Local Comm Utilization

• Generally follow the global throughput curves.
• It is directly related to frequency of messages
  between MDAS and local DB sites.

32
Simulation observation

• Global throughput is not the only issue.
• Completion rate is also an important measure.

33
Simulation conclusion

• V-Lock have best completion rate as long as the
  global concurrency degree is not too high. And it
  has the best global throughput when global
  concurrency degree is 7 or lower.
• Site Graph leads to good global throughput, at
  the expense of completion rate, as concurrency
  degree increases.
• Potential Conflict Graph (better) and Force
  Conflict result in a relatively steady global
  throughput.
• Potential Conflict Graph and Force Conflict have
  fair completion rates between Site graph and
  V-Lock.

34
Resubmission

• Effect of resubmission of a failed transaction.
• Observation
• Low completion rates as the global
  concurrency degree increases.
• Attempt
• Give more (2,3,6,and 10) chances to an aborted
  transaction.
• Expectation
• Higher completion rate and higher global
  throughput.

35
Resubmission Completion rate
36
Resubmission Completion rate

• Effect on completion rate
• Resubmission leads to higher completion rate for
  all four protocols.
• Most significant for Site Graph.
• Least significant for V-Lock.

37
Resubmission Response time
38
Resubmission Response time

• Effect on response time
• Resubmission leads to longer response time for
  all four protocols.
• Significant to Site Graph because it has the
  lowest completion rate.

39
Resubmission Global throughput
40
Resubmission Global throughput

• Effect on global throughput
• Resubmission does not change much on V-Locks
  global throughput
• Resubmission leads to lower global throughput for
  all other 3 protocols.

41
Resubmission

• Effect of resubmission of a failed transaction.
• Result
• Higher completion rate
• Longer response time.
• Similar(V-Lock) or lower global throughput.
• Conclusion
• Resubmission increase completion rate but leads
  to a similar or lower global throughput.
• Resubmission is not recommended unless a higher
  completion rate is desired.

42
Current Work

• Survey on mobile querying/transaction in wireless
  environment.
• The survey will be finished in 3 weeks.

43
Future Work

• Simulate with non-busy global/local systems.
• Considerate weak connection and disconnection.
• Add a small cache to the Mobile Unit.
• Choose/design the suitable cache policy. (This
  semester)
• Enable mobile Databases. (replica, update, etc.)
• Design a concurrency control protocol for mobile
  transaction for wireless environment.

44

• Questions and comments??