Agent architecture for mesh based simulation systems

1
Agent architecture for meshbased simulation
systems

• Krzysztof Banas
• Institute of Computer Modelling
• Cracow University of Technology

2
Outline

• Architecture for sequential mesh based PDE
  simulation codes
• Classical approach to parallelization
• Current trends in distributed software systems'
  architecture
• Agent architecture for PDE simulation codes

3
Modular architecture for PDE simulation codes

• Fundamental computational kernel modules
• mesh manipulation
• field data
• linear equations solver
• (optional) linear solver interface and/or adapter
• Complex interactions among modules
• An example for finite element codes

4
(No Transcript)
5
Classical approach to parallelization

• Domain decomposition
• embarrassingly parallel integration and assembly
• overlapping or non-overlapping iterative linear
  solver
• Message passing
• Centralized management
• the same decomposition for all modules
• the same global address space mechanisms
• An example for finite element codes

6
(No Transcript)
7
Current trends in distributed systems software
architecture

• Loosely coupled components
• Dynamically configurable systems
• Advantages
• more flexibility
• better utilization of dynamically changing
  resources
• Disadvantages
• lower performance for static configurations
• An example for finite element codes

8
(No Transcript)
9
Agent architecture for mesh based simulation
systems

• The need for configuration optimization
• adaptation to different computing environments
• multi-core processors
• clusters
• grids
• adaptation to different solution strategies
• higher order approximation
• multigrid solvers

10
Agent architecture for mesh based simulation
systems

• Agents as a mean to optimize complex simulation
  systems
• the knowledge of possible hardware and middleware
  configurations
• the knowledge of simulation software requirements
• Separate from the simulation code
• Acting on its behalf

11
(No Transcript)