Workflow Management with Service Quality Guarantees

1
Workflow Management with Service Quality
Guarantees

• Paper by Gillmann, Wiekum Wonner
• Presented by Samir Tartir
• March 18, 2004

2
Talking Points

• Introduction
• Architecture
• Model
• Extended Model
• Evaluation
• Experiment Results

3
Motivation

• Current web services have performance problems,
  mostly caused by servers.
• Servers need scarce and expensive human experts.
• Business requirements change frequently.

4
Solution Workflow Management

• Can be defined explicitly in the middle-tier, or
  implicitly in business processes.
• Examples of WFMS MQ Workflow, Staffware,
  Biztalk, E-Speak.
• WFMS can be placed on multiple servers to enhance
  performance.

5
Questions

• How to devise a strategy for the degree of
  replication?
• How to configure the entire system such that it
  can guarantee the quality of service.

6
Answer Goliat

• Stands for Goal-driven auto-configuration tool.
• Automates configuration of a distributed WFMS.

7
Architecture
8
Interface - 1
9
Interface - 2
10
Goliat Output

• Total load.
• Maximum sustainable throughput.
• Expected mean waiting time at each server.
• Mean turnaround time
• Expected availability.
• Expected performability.

11
Stochastic Modeling of Workflow Behavior
Example EP
12
The Flow Process

• In CTMC the probability of entering the next
  state within a certain time depends only on the
  currently entered state.
• P (Pi,j)
• H (Hi)
• S Si I 0 n-1

13
Definition

• Ft(T) Zt U SA, whereFt(T) the flow process
  of workflow type tZt the set of possible
  execution states of the workflow.SA an
  absorbing state.
• Three conditions.

14
EP in CTMC
15
Mean Turnaround Time

• The MTT Rt is the mean time the CTMC needs to
  enter the absorbing state.
• Calculated using methods like Gauss-Seidel
  algorithm.

16
Induced Load Per Workflow Instance

• L (Lxa) Load Matrix Service Requests on server
  type x by executing a single instance of activity
  a.

17
Performance Model

• Service types are uniformly distributed across
  all servers of the same type.
• Each server type is considered to be a
  well-configured building block.

18
Performability Model

• A state in CTMC is a k-tuple with k being the
  number of different server types with the WFMS
  and each entry of the tuple represents the number
  Xx of the currently available servers of server
  type x when the CTMC is in that state.
• Rates depend on MTTF and MTTR of server types

19
Extended Performability Models
Modeling of Control Flow Loops

• Loop properties1. One entry point.2. One exit
  point.3. Has a positive lower bound. (?)4. Has
  a finite upper bound. (?)

20
Modeling of Control Flow Loops - 2

• State expansion Each state Sx ? Si,,Sij is
  replaced by cloning to be Sx,1,,Sx,n.
• Entering add a transition between Sx to Sy,1.
• Next add a transition between Sij,o to Si,o1.
• Internal add a transition between Sx,o to Sx,j.
• Exit add a transition between Sij,n to Sx.

21
Evaluation
Testbed

• WF engine is Mentor-Lite.
• Load generator.
• Monitoring module.
• Activity module.
• Crash Module.

22
Parameter Setting

• Mean turnaround time See table.
• Loops 1 to 3 iterations.
• CC Payment 50.
• CC Failures 10.
• Service times See table.
• Downtimes See table.

23
Results

• See paper.