FLAME : A parallel agent

1
FLAME A parallel agent based framework using
X-machines
Co-funded by the European Commission within the
Sixth Framework Programme
Website www.flame.ac.uk
Introduction
X-machines are finite state machines with the
inclusion of memory which influences the state
transitions in the model. They have been used to
specify and test software systems and are also
being used for modelling more complex structures
such as agents in agent based models. FLAME uses
this paradigm accompanied with abilities to
parallelize the models allowing high
concentrations of agents with more complex
structures to be simulated in finite time. Two
examples from the fields of biology and economics
have been described below as case studies.
Chris Greenough David Worth Lee Shawn
Chin Rutherford Appleton Laboratory, UK
Mariam Kiran Simon Coakley Mike
Holcombe Department of Computer
Science, University of Sheffield, Sheffield, UK
Modelling
Biology Keratinocyte cell model
Economics Labour market model
State transition diagrams for two agents - firm
and household.
State transition diagram for a cell showing the
different forms it can exist in.
Structure of a X- machine agent
Following from the input/output messages the
function dependencies can be created. These allow
how the different modules can be parallelised
over a set of processors.
The dotted arrows represent data dependencies
between the functions. These represent the
synchronisation points which insure that all
functions prior to that point have finished
processing. These keeps track of all functions to
be synchronised when running on multiple
processors.
3 synchronisation points
4 synchronisation points
FLAMEs Layer Structure
Current works

• Mano 1024 nodes of dual-core 700MHz PowerPC
  chips.
• Hapu 128 x 2.4GHz Opteron cores, with 2Gb
  memory per core.
• NW_GRID 32 SUN x 4100 nodes. Each node
  contains 2 Dual Core 2.4Ghz Opterons with 8GB of
  memory. That brings the total processor count to
  192 Dual-Core Opterons (384 processor cores).
• HPCx Total of 2560 processors.

Few Results
Acknowledgements
We would like to acknowledge the works of Neil
Walkinshaw and Phil McMinn in contributing to the
modelling methodologies.