Meeting: Applications on Grid
26 Marzo 2007
Paolo Arena, Maide Bucolo, Arturo Buscarino, Riccardo Caponetto,
Federica Di Grazia, Giovanni Dongola, Luigi Fortuna, Mattia Frasca,
Davide Lombardo, Luca Patanè, Francesca Sapuppo
Dipartimento di Ingegneria Elettrica Elettronica e dei Sistemi
Università degli Studi di Catania, Catania, Italy
Outline
Research fields
• Analysis and control of high dimensional complex
systems (Prof. Fortuna, Prof. Arena)
• Massive signal processing in neuroengineering
(Prof. Bucolo)
• Computational Fluido-Dynamics and microfluidic
system control (Prof. Bucolo)
• Distributed Genetic Algorithms (Prof. Caponnetto)
• Air pollution forecast (Prof. Caponnetto)
3D network of coupled dynamical systems
Analysis and control of high dimensional complex
systems
– 3D network of coupled dynamical systems: nonlinear
oscillators (Prof. Fortuna)
– 3D network of coupled dynamical systems: chaotic
oscillators (Rossler, Lorenz, Chua, …) (Prof. Fortuna)
– 3D network of coupled dynamical systems: neuron models
(Izhikevich, HR, …) (Prof. Fortuna)
– Simulation of high dimensional neural networks: Liquid
State Machine (Prof. Arena)
– Simulation of high dimensional neural networks: modeling
mechanisms of the brain (Prof. Arena)
3D network of coupled dynamical systems
Nonlinear oscillator
FitzHughNagumo
60x60x60 cells
vb
 1

u

u
(
1

u
)(
u

)  D 2u

a
 
v  u  v
3D network of coupled dynamical systems
Chaotic oscillators
Rossler
60x60x60 cells
Lorenz
60x60x60 cells
Chua
80x80x80 cells
 x   y  z  D 2 x

 y  x  ay
 z  b  xz  cz

 x   ( y  x)  D 2 x

 y  rx  y  xz
 z  xy  bz

 x1   ( x2  f ( x1 ))  D 2 x1

 x2  x1  x2  x3
 x    x
2
 3
3D network of coupled dynamical systems
Neuron models
Izhikevich
30x30x30
neurons
Inferior Olive
50x50x50
neurons
HR neuron
50x50x50
neurons
 dv
2

0
.
04
v
 5v  140  u  I

 dt

 du  a (bv  u )

 dt
x ( x   )(1  x )  y  I
 dx
 D 2 x
 dt 


 dy
 z  r ( A  z 2  r 2 )

dt

 y

 dz
2
2
r    x
 dt  z  z ( A  z  r )
M


 x  y  ax 3  bx 2  D 2 x

2
 y  c  dx  y
 z  r ( s ( x  x )  z )
c

Simulation of high dimensional neural networks:
Liquid State Machine
input(t)
• Non regular high dimensional neural
networks processing continuous stream of data
• Universal power computation
• Learning algorithms train a readout map to
attribute a meaning from the Liquid status
dv
 0.04v 2  5v  140  u  I
dt
du
 a (bv  u )
dt
Simulation of high dimensional neural networks:
modeling mechanisms of the brain
“Simple” insect brain is composed by very
complex and highly connected neural
structure made of hundreds of cells
Environment for the Biopotential study
1 file1 Area
2 file2 Areas
MEG
1 subject
SX/DX
•Sample frequency =254.31 Hz
•Exercise period =51 min
•N° of channels= 148
•Codification =32 bit
•Data sets= 6
≈ 2.7 GB per subject
•Sample frequency =251 Hz
•Exercise period =30 min
•N° of channels= 37 per hemiphere
•Codification =32 bit
•Data sets≈8
≈ 100 GB
STATISTICAL ANALYSES
≈ 2 GB per subject
Environment for the Biopotential study
Geometry and Type of data.
PRE-PROCESSING: Statistical analysis,
Frequency & Saturation filtering
PROCESSING: Temporal ICA, Spatial
ICA,Power Distribution, Nonlinear analysis.
Plot 1D, 2D, 3D.
Environment for the Biopotential study
Environment for the Biopotential study
Data Storage (raw data)
Computational
Complexity
Spatial
Area, channel
Parallel Data Processing
Temporal
Minute, second, msecond
Data Storage (result)
Comparison: Qualitative versus Quantitative
Research: Biopotential Study via Image processing Representation
Clinic: Data Base for Diagnostics Support
CFD & μ-Fluidic System control
In Vitro
µFluidic Devices for
Laboratory Analysis
Lab-on-Chip (LOC)
LOC
In Vivo
µFluidic Systems
Microcirculation
MEMS
IC
Arteriole
(15-100 m)
100-600μm Section
Serpentine Mixer
Splitter
Venule
(10-80 m)
Capillaries (6-8 m)
CFD & μ-Fluidic System control
Control
System
Monitoring
System
Microfluidic
Process
Methodological
Framework
Opto-Sensing
System
Experimental
Workbench
CFD & μ-Fluidic System control
Navier-Stokes Equation
water
air
 

 

 u

 u  u      w u  u t  p   w g  F
 t

w 
 u

 u  u      a u  u t  p   a g  F
 t

a 


 u

   u  u      u  u t   p   N     g  F , u  0
 t

   w   a   w H  
   w   a   w H  
Multiphysics
Fluido-Dynamics
Eq. Navier-Stokes





u

  2u   u  u  p  F
t
with

u  0
 ts
Level–Set Equation

c
  Dc   R  u c  0
t
Microfluidic
Electromagnetisms

 u    0
t
Chemistry
Eq. Reaction
B
 K on c Rt  B   K off B  0
t
Eq. Laplace
 V   0
with

E  V


D   0 r E
Mechanics
Eq. Convection-Diffusion
Pump1
Systems
T
 ts C
t
Thermodynamics

  kT   Cu T  Q
Eq.Continuity
Eq. Joule Effect
H2O
f1

Q  E 2
Pump2
Air
f2
CFD & μ-Fluidic System control
Two phase
fluid
Multi-phase
Fluid
Single
Complex
channel
network
2D
3D
Distributed Genetic Algorithms
C++
Multi objective optimization problems
Genetic Algorithms
Multi objective optimization
Parallel Processing
problems
Design variables
X  x1 , x2, ..., xn 
Objective function F   f1 ( X ), f 2 ( X ),..., f m ( X )
Divided Range Genetic
Algorithms (DRGAs)
Constraints
G j ( X )  0 ( i = 1, 2, … , k)
Parallel Genetic Algorithms can be classified in three main types :
Global Master-Slave
Island or Coarse-grained or Distributed Model
Hybrid or Hierarchical Models
Global Master-Slave Model
C++
Slave
Slave
Compute fitness
Compute fitness
Slave
Compute fitness
Slave
MASTER
Slave
Compute fitness
Distribute
individuals
Compute fitness
Slave
Slave
Compute fitness
Slave
Compute fitness
Compute fitness
Island or Coarse-grained or Distributed Model
C++
Subpopulation
Subpopulation
Subpopulation
Subpopulation
Migration
process
Subpopulation
Subpopulation
Subpopulation
Subpopulation
Hybrid Models (Island + Master-Slave)
Subpopulation
Subpopulation
Subpopulation
Subpopulation
C++
Hybrid Models (Island + Island)
Subpopulation
Subpopulation
Subpopulation
Subpopulation
C++
Divided Range Genetic Algorithms
C++
F2
better
F1
Divided Range Genetic Algorithms
C++
F2
better
F1
Divided Range Genetic Algorithms
C++
 are suitable for the parallel processing.
 can derive the solutions with short time.
 can derive the solutions that have high accuracy.
 can sometimes derive the better solutions compared
to the single island model.
GRID and its application for air pollution forecast
C++
 Modelling of air
polluted
atmosphere by
parallel
applications
 Potential users
 environmental
authorities
 local
governments
 media
Towards the GRID
• 3D networks of coupled dynamical systems: E^3
– The tool has to be adapted to Unix
– Server communication has to be extended to the
GRID case
• Massive signal processing in neuroengineering:
Matlab
– The environment has to be ported in Scilab or C
language
• CFD & μ-Fluidic System control
– Finite Elements Method
Towards the GRID
• 3D networks of coupled dynamical systems: E^3
and Matlab
– The tool has to be adapted to Unix
– Server communication has to be extended to the GRID
case
– Can multiserver feature be supported?
• Massive signal processing in neuroengineering:
Matlab
– The environment has to be ported in Scilab or C
language
• CFD & μ-Fluidic System control
– Finite Elements methods
E^3: an Emulator for Complex Systems
• Dynamical systems (represented by a
set of ordinary differential equations)
• Interactions among many elementary
units to constitute a complex 3D
structure
• Interactions among 3D structures to
constitute a complex system
• E^3: an universal emulator for
complex systems
E^3: an Emulator for Complex Systems
• A complex system can be defined a system
made of interacting nonlinear units (dynamical
systems)
• Characteristics of E^3
–
–
–
–
capabilities for parallel computing
single cell dynamics can be defined by the user
network connections may be defined by the user
procedures for output visualization and elaboration
=
E^3: Parallel computing
• E^3 is divided in two main modules:
– Client (definition of the complex systems to
be simulated)
– Server (simulation of a given number of cells)
E^3: strategies for parallel computing
Server 1
Server 3
Server 2
Server 4
• Domain decomposition: single program
multiple data
• Client acting as Master process
• Communication method: message passing
E^3: Software implementation
GNU GSL
COMMAND INTERF.
VISUALIZATION
NUMERIC ENGINE
Mathematica
Matlab
CLIENT GRID ENGINE
SERVER GRID ENGINE
RPC
Java RMI
E^3 Kernel API