Quality Aware Service Planning in Computational Grids

1
Quality Aware Service Planning in Computational
Grids
Sharath Babu MusunooriNetworks and Distributed
SystemsSimula Research Laboratory, Norway2nd
International Doctoral Symposium on Middleware
05 28 November 2 December, 2005 Grenoble,
France
2
Motivation

• Grids are future platforms for parallel
  distributed systems
• Grid applications acquire required network and
  computational resources of the grid environment
• Acceptable application performance in a grid
  environment remains a difficult engineering
  challenge
• In particular, in the case of quality-senstive
  applications
• Quality-aware service configuration
• A mapping of individual service components of a
  given application onto the underlying grid
  resources while satisfying the specified quality
  requirements.

3
Presentation Outline

• Background
• Problem description
• Quality-aware service planning
• Quality deviation model
• Introduction to Learning Automata (LA)
• Modelling LA
• Learning algorithm
• Simulation
• Conclusion

4
Background (1)

• Application
• A service graph with corresponding communication
  links
• A service is an independent functional component
  blueprint
• Grid Environment
• Each node hosts a middleware agent (capsule)
  representing a local run-time environment
• A capsule includes functionality to organize
  local resources (CPU, mem), provides basic
  life-cycle management support for application
  services

5
Background (2)

• Component architectures simplify application
  development
• The notion of independent composition and
  deployment
• A clean separation between logical and physical
  layers of component-based programming
• Platform-managed QoS approach
• Support quality characteristics of the
  application during deployment and runtime

6
Problem Description (1)

• Component technology simplifes application
  design
• Service quality depends on amount of resource
  availability
• Capsules are containers for hosting services
• How to partition service set into the number of
  capsules sub-sets such that all services get
  sufficient resources to allow them to achieve at
  least the minimum quality

7
Problem Description(2)

• The objective of a service is to maximize its
  quality
• The optimal service configuration
• Simultaneously tries to maximize the quality
  objective of all services from the given set of
  avaialble resources
• The combinatorial explosion makes it a complex
  task
• Seek a feasible solution instead of optimal one.
• A feasible service configuration
• Tries to achieve at least the minimum required
  quality levels for all services from the given
  set of available resources

8
Quality-Aware Service Planning (1)
9
Quality-Aware Service Planning (2)

• General weighted sum solution technique
• Derives a solution set by finding the convex sum
  of objectives
• User chooses appropriate weights
• solutions points are pulled over one or more
  objectives
• Even the optimal configuration may not guarantee
  the overall application performance

10
Quality Deviation Model (1)

• Finds a nice compromise between the quality
  objectives
• Aspired quality levels allow the service user to
  set more realistic quality targets
• Error is a distance between the aspired and
  actual quality levels
• Proposes to minimize the error
• Fine-grained trade-off
• Guarantees the overall application performance

11
Quality Deviation Model (2)

• Estimated service configuration when grid
  environment is over provisioned with excess
  resources

12
Quality Deviation Model (3)

• Estimated service configuration when resources
  of the grid environment are scarce

13
Introduction to Learning Automata (1)

• A simple model for adaptive decision making in
  unknown random environments
• Example a person finding the best route from
  home to office.

14
Introduction to Learning Automata (2)

• The automaton adapts itself to the environment
• Complete generality of the concept of the action
  set
• LA is a simple unit from which complex systems
  could be constructed
• Convergence results exist
• A general approach to stochastic optimization
• Fixed-structure stochastic automata (FSSA)

15
Modelling LA in the Context

• Each service is an automaton
• Grid platform is an environment
• Services choose capsules and learn from the grid
  environments feedback

16
Learning Algorithm (1)

• Establishes autonomous learning environment for
  services
• Unsatisfied services move between capsules
• Move is a model of service move in computing
• Services learn and even unlearn from the grid
  environment
• Learning factor is a service confidence level
• A positive feedback allows services to increase
  their confidence
• Services with zero confidence adjust with their
  minimun required quality levels

17
Learning Algorithm (2)

• Always finds a feasible configuration
• In best case of grid platform with excess
  resources
• Probability of any service get reward is high
• All services gain their maximum quality
• In the worst case of grid platform with scarce
  resource
• Probability of any service get penalized is high
• All services get satisfied with at least minimum
  specified quality

18
Simulation
19
Conclusion Future Work

• Quality deviation model provides a fine-grained
  trade-off
• First learning algorithm designed and
  implemented
• LA establishes an autonomous learning
  environment
• Learning algorithm always converges to a
  feasible configuration, scales well with number
  of services
• Quality Synchronization
• Different variants of learning
• Application service reconfiguration

20
Questions
21
Learning Algorithm

1. repeat
2. Select a capsule, Cm with least quality
3. Select a service, Si with least quality on
   capsule Cm
4. Move the service, Si to a random capsule, Cn
5. if (move is successful)
6. Reward all services on capsule Cm
7. Reward the service, Si on capsule Cn
8. else
9. Penalize all services on capsule Cn
10. All services with zero confidence adjust
    with minimum quality
11. until (all services are happy)

22
Simulation

• Initialization
• Generated grid platform confirms the generality
  of the physical grid
• Minimum service quality requirements are set
• Makes sure that there is a feasible
  configuration
• Implemented in C and ran on Condor system
• Tested for various combination of services and
  capsules
• 500 simulations for each combination
• Efficiency is measured in terms of the number of
  service moves