Transcript glogin - A Multifunctional Interactive Tunnel into the Grid

Making the Best of Your Data Offloading Visualization Tasks onto
the Grid
Semiautomatic Generation of Transfer Functions
through Grid-based Parameter Studies*
Peter Praxmarer
GUP, Joh. Kepler University Linz
[email protected]
*Partially funded by the Austrian Grid Project
(BMBWK GZ 4003/2-VI/4c/2004)
Agenda
•
•
•
•
•
•
What are Transfer functions?
Goal
Approach
Characteristics
Use of the Grid
Conclusion and Future Work
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
2
What are Transfer
Functions?
• Volumetric data represents:
– A volume with some scalar property at every point
– Properties can be: density, temperature, chemistry, …
• Rendering methods:
– Isosurface: A 3d contour is created at a selected
density, the resulting surface shows all the regions
that are more (or less) dense than the chosen contour
level.
– Raytracing/Raycasting: A transfer function provides a
mapping of the density values to color and
transparency.
• Thus: D->(R,G,B,A)
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
3
What are Transfer
Functions?
• Example:
– Given: Some volume data, a transfer function for
transparency only.
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
4
What are Transfer
Functions?
• Example:
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
5
What are Transfer
Functions?
• Example:
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
6
What are Transfer
Functions?
• Example: Adding a color map
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
7
What are Transfer
Functions?
• Example: Adding a color map
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
8
Task
• Find transfer functions for visualizing volume
data
• Application:
Offline-Rendering (Raytracing) of gas
distributions in galaxy clusters
• Given:
– Voxel data (density, temperature)
• Wanted:
– Mapping: D->(R,G,B,A)
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
9
Properties of the data
• Astrophysical volume data generated by N-body SPH
simulations
• Unlike MRI data the astrophysical data is highly
amorphic
– In MRI data there are sharp boundaries -> we can use the
gradient information for detecting boundaries
– Galaxies often consist of a lot of gas which gradually gets denser
-> the gradient alone is not sufficient to generate good visual
representations (especially with isosurface rendering)
• Finding a transfer function that reveals the ‘interesting’
parts of the simulated data is difficult
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
10
Approach
•
Apply a genetic algorithm:
1. Generate an initial population of transfer
functions
2. Evaluate the fitness of each chromosome
3. Select the best chromosomes (transfer
functions) for the next population pi+1
4. Recombine the chromosomes
5. Mutate some chromosomes
6. Perform steps 2 to 6 until a good-enough
transfer function has been found.
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
11
Encoding (1)
• A chromosome represents one transfer
function
• Each chromosome has N<50 genes
• Each gene stores
– Density value
– Color
• Thus: Mapping D -> (R,G,B,A)
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
12
Encoding (2)
Chromosome (transfer function)
Gene 0
Density value
Hue,
Saturation,
Value
Talk at CGW05
Gene 1
Density value
Hue,
Saturation,
Value
…
Gene N-2
Gene N-1
Density value
Hue,
Saturation,
Value
Density value
Hue,
Saturation,
Value
Peter Praxmarer,
GUP, Universität Linz
13
Generating an initial
population
• Collecting statistical measures:
– Histogram of the voxel data
– Mean, average, mode
• Used to generate the initial population
using a heuristics.
• Selection based on
– Density interval
– Frequency
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
14
Evaluation of the
fitness
1. Render the population on the grid
2. Present the resulting images to the user
3. The user judges the transfer functions
according to a like/don’t like scheme
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
15
Selecting the next
population
• Select transfer functions for popi+1
proportional to their fitness (the better a
transfer function is, the more often it is
selected)
• Introduces a bias towards ‘better’ transfer
functions
• Is not sufficient to generate new transfer
functions
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
16
Recombination (1)
• Allows a transfer function to move towards
interesting ‘places’
• Applied with a probablity pcrossover (usually
pcrossover = 0.7)
• Generates two offsprings from two parent
chromosomes
• The parent chromosomes are chosen by
random from the previously selected
chromosomes
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
17
Recombination (2)
For example:
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
18
Mutation
• Applied with probability pmutation (usually
pmutation=0,001) per gene
• Randomly change the gene (color or
density value)
• Introduces new solutions into the search
space
• Prevents the premature convergence to
local optima
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
19
Application
characteristics
• The user directs the search
– Only the domain expert knows what he wants
to see
• Allows finding transfer functions for
volume data with an amorphous structure
(galaxy data vs. MRT data)
• Requires large computational power to
render the images of a population
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
20
Rendering on the Grid
• Use today’s grid technology to distribute
the load on various resources
• Prerequisites:
– POVRay
– Grid infrastructure
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
21
Components
• Consists of
– GUI: Presents the rendered images
– Master: Server that maintains the connection to
clients; Runs on the same machine as the GUI.
– Clients: Are running on the Grid. Connect back to the
Master and receive and execute commands from the
master:
•
•
•
•
Retrieve Density data
Retrieve Scene description
Execute Rendering
Send Output back to GUI
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
22
Parameter study
• The transfer function is saved as a
colormap in the POVray scene
description file
• Clients receive commands to
execute the Rendering and
transfer back the results
• Rendering is parallelized across
multiple Grid nodes
• Data transfer using GridFTP
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
23
Results (1)
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
24
Results (2)
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
25
Results (3)
• Rendering time: ~60 sec / transfer function
with POVRay; resolution 600x600
• By Parallelizing the POVRay rendering the
rendering time can be significantly
reduced (depends of number of available
nodes)
• Population size: typically 16, up to 64
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
26
Conclusion
• Supports astrophysicists in finding useful
transfer functions for visualizing their simulated
data
• The astrophysicists directs the search to what he
wants to see in his data
• Due to the use of Grid Technology he is able to
explore many different settings at once in a
considerably short time
• Due to ray-tracing he gets high-quality
representations of the volume data
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
27
Future Work
• Improve heuristics for generating the initial
population
• Use good transfer functions from astrophysicists
as a starting point
• Improve the GUI to allow manually changing the
presented transfer function. This should be done
locally on the workstation to provide interactivity.
• Reduce the response time
• …
Talk at CGW05
Peter Praxmarer,
GUP, Universität Linz
28