Towards WorkflowDriven Database System Workload Modeling

1
Towards Workflow-Driven Database System Workload
Modeling

• Du Naiqiao, Ye Xiaojun and Wang Jianmin

2
Outline

• Introduction
• Workload model
• Model with Markov process
• Model with Petri-net
• Conclusion

• Why we need workflow workload?
• How to model it?

3
Two main questions

• Why we need workflow workload?
• Workflow patterns exist in real system.
• Workflow workload will affect the performance.
• How to model it?
• Markov process
• Petri net

4
Introduction-motivation

• Existing benchmarks
• Transaction workload in TPC-C and TPC-App are
  treated independence, only keeping the mixing
  ratio
• TPC-E dependence between transactions
• Intuitive e.g. E-Commerce system
• search -gt select -gt buy
• Data mining area customer access pattern exists
  in OLTP system
• Will such patterns affect the database
  performance? HOW?

5
Introduction-to do what

• Relationship between transactions should be
  considered in OLTP performance test.
• The relationship here is different from TPC-E
• Weak relationship----random transaction sequence
  comply with some kind of access pattern.
• Strong relationship----the sequence of
  transactions is determined obviously.
• Target
• Model the relationship between the transactions
• Keep the transaction mixing ratio (traditional
  benchmark requirement)
• AT THE SAME TIME.

6
Introduction-Related work(1)

• Menasce et al, A Methodology for Workload
  Characterization of E-commerce Sites, ACM
  E-Commerce Conf 99
• Customer Behavior Model Graph (CBMG) describe
  the behavior of groups of customers who exhibit
  similar navigational patterns.
• Show the steps required to obtain parameters (the
  arrival rate of session initiation requests and
  the average server-side think time).
• Propose a clustering algorithm (k-means) to
  characterize workloads of e-commerce sites in
  terms of CBMGs.

7
Introduction-Related work(2)
8
Introduction-Related work(3)

• Yao et al, Mining and Modeling Database User
  Access Patterns, ISMIS 2006
• Kth-order Markov process to model database users
  behavior.
• Data mining techniques to mine database users
  access patterns
• pruned Markov models the states and state
  transitions are pruned according to the state
  support threshold (i.e., the minimal support
  value for the node) and the confidence threshold
  (i.e., the minimal confidence value for the edge)
  ,the number of states and the number of state
  transitions are reduced.

9
Introduction-Related work(4)
10
Two main questions

• Why we need workflow workload?
• v Workflow patterns exist in real system.
• Workflow workload will affect the performance.
• How to model it?
• Markov process
• Petri net

11
Mapping between workflow workload and transaction
workload
12
Model with Markov process(1)

• Transaction can be activated by controller or
  other transactions.
• Every transaction is a state one-order Markov
  process.
• State space S1,2n
• Matrix Ppij, probability that transaction j
  activates directly after transaction i

13
Model with Markov process(2)
14
Model with Markov process(3)
15
Two main questions

• Why we need workflow workload?
• v Workflow patterns exist in real system.
• Workflow workload will affect the performance.
• How to model it?
• v Markov process
• Petri net

16
Model with Petri-net(1)

• Model TPC-C or TPC-App with Petri-net
• PaPbPcPd1
• One Addition condition end or not

17
Model with Petri-net(2)
A,B,C and D can be replaced by Subflows. The
benefit of using Petri-net is the ability to
model parallel situations. a c b d
18
Model with Petri-net(3)
Define is the expectant occurrence
times for T1 in flow A.
19
Model with Petri-net(4)
20
Model with Petri-net(5)
21
Two main questions

• Why we need workflow workload?
• v Workflow patterns exist in real system.
• Workflow workload will affect the performance.
• v How to model it?
• v Markov process
• v Petri net

22
One question remains

• How will the workflow workload (customer access
  patterns) affect the performance?
• Experiment by TPC-App
• As the effect of web server is not determined,
  the results are not included in this paper. But
  we believe the web server is not the bottle neck.

23
Two main questions

• v Why we need workflow workload?
• v Workflow patterns exist in real system.
• v Workflow workload will affect the performance.
• v How to model it?
• v Markov process
• v Petri net

24
Conclusion

• Transaction relationships should be considered.
• Two different methods are used to model the
  relationship between transactions in OLTP
  systems.
• The mixing ratio for transactions is studied,
  which will be benefit for designing or modifying
  the OLTP performance test benchmark.
• Future work
• More work on how customer access patterns will
  affect the database performance.
• Analysis on the parameters dependence between
  related transactions
• More complicated models, e.g. model the
  transaction relationships with addition Petri-net
  patterns.

25

• Thank you