Autonomic Execution of Web Service Compositions Cesare Pautasso, Thomas Heinis, and Gustavo Alonso

1
Autonomic Execution of Web Service
CompositionsCesare Pautasso, Thomas Heinis, and
Gustavo Alonso

• Koji Hashimoto
• 2009/02/26

2
Background

• JOpera project (www.jopera.org)
• Eclipse-based tools for Web service composition
• Workflow-based language, execution, monitoring,
  and debugging
• JOpera distributed engine
• Run on cluster of computers to handle large
  workloads
• Autonomic computing feature

3
About JOpera
Process
Task Web service operation
Task Web service operation
Screenshot taken from http//www.jopera.org/node/8
4
JOpera Distributed Engine Architecture
5
Autonomic Controller Algorithm

• Monitoring Phase
• Monitor the system status based on Information
  policy
• Planning Phase
• Plan a configuration change based on Optimization
  policy
• Third Phase
• Carry out an actual configuration change based on
  Selection policy

6
Autonomic Controller Policies

• Information Policy
• Define which performance indicators should be
  monitored
• Queue length (Process queue Task queue)
• Resource utilization (how long the system uses a
  cluster node)
• Optimization Policy
• Define the mapping of performance indicators to
  actions
• Trade-off between performance and resource
  utilization
• Simple Threshold Policy
• Differential Policy
• Proportional Policy
• Selection Policy
• Define how to select actions the system is going
  to take

7
Simple Threshold Policy

• Start one thread if Queue Length gt T
• Stop one thread if Queue Length 0
• No action if 0 lt Queue Length lt T

8
Differential Policy

• Queue Length Variation ?q q(t) q(t-1)
• Start one thread if ?q gt Tstart
• Stop one thread if ?q lt Tstop

9
Proportional Policy

• Start/stop one or more threadsproportionally to
  ?q

10
Experimental Evaluation

• Workload
• 800 concurrent identical processes
• Each process is composed of 10 parallel tasks
  with 8 sec. invocation time
• 32 nodes (one Navigator/Dispatcher per node)
• Autonomic controller algorithm runs every second.

11
Simple Threshold Policy

• The size of task queue grows high because
  Navigators and Dispatchers are balanced.

12
Differential Policy

• of Dispatchers follows small increase of Task
  queue length.

How do threads decrease when the queue length 0?
13
Proportional Policy

• Adapt faster to the workload peak

How do threads decrease when the queue length 0?
14
Comparison of the policies
15
Conclusions

• JOpera is an Eclipse-based Web service
  composition tool.
• Three autonomic control policies are proposed and
  evaluated.

16
Critiques

• JOpera is available and widely used.
• The autonomic policies were implemented and
  evaluated.
• Resource utilization may not be evaluated
  correctly.
• Dispatcher ? Service provider
• Where are Service providers located?
• More experimental evaluation is necessary(the
  authors mentioned though).
• Different process structure
• Different threshold setting
• Multiple threads per node

17
Extensions

• Load balancing
• Dynamic threshold setting strategy

18
SASSY vs. JOpera

• SASSY

• JOpera

• Workflow-based language QoS annotation
• Distributed coordination
• Self-Adaptation
• performance
• security
• reliability/availability
• etc.
• Work in progress

• Workflow-based language
• Centralized coordination
• Autonomic computing
• performance
• resource utilization
• Available and widely used