GPU-Based Interactive Visualization of Billion-Point Cosmological Simulations Tamas Szalay,

GPU-Based Interactive
Visualization of Billion-Point
Cosmological Simulations
Tamas Szalay,
Volker Springel,
Gerard Lemson
The Visualization Problem
• Getting better at storage and processing
– Distributed databases, clouds, etc…
– I/O needs to be only as fast as computation
• Doesn’t work for visualization
– Would need to read all the data every frame or have it
all in memory
– Even rendering itself would be prohibitive
• Could use pre-rendered movies
– Trial and error takes time
The Aquarius Simulations
A series of n-body dark matter simulations
Run from the early universe to today
Box roughly the size of galactic neighborhood
Run five times at different particle resolutions
– Lowest has 2.3 million, takes up about 25 GB total
– Highest has 4 billion, and takes up 20 TB
• Each version has point data in 128 ‘snapshots’,
with positions and velocities
• Movies have been rendered, took weeks
Visualization Motivation
• Certain types of analysis very difficult otherwise
– Qualitative impressions of gravitational structures
• Verification of simulation and accuracy of
structure finding
• Identification of events of interest
– Comparisons of multiple objects
– Two colliding gravitational clusters
– Dark matter streams
• Public outreach
Hierarchical Rendering
• Don’t need to render everything
– Saturates screen anyway
• So show the same data, but render less
– Create different levels-of-detail for entire dataset
– Load different parts from different levels as needed
• Put levels-of-detail on fast storage system
• And give it a rendering front-end
• Think Google Microsoft Maps
Level Structure
• Chose spatial octree because it is simple and general
• Each node also has associated data
– All of the data spatially contained within the cube
– Except simplified to <= N points
• Deeper in the tree means higher resolution
• Organized this way on disk
Selective Loading
• What resolution data to load in what spatial
– Close to viewer in high detail and far away in low
• Can use the on-screen size of the relevant
octree cube to determine resolution
– Means, in theory, visually equivalent to entire dataset
• Automatically scales to rendering hardware
• Can spread out through time as well
Rendering Front-End
• GPUs are fast
– Really fast
– Can do an unbelievable amount of computation in
rendering pipeline
– Allows tool to still do significant processing
• The actual rendering algorithm:
– Brightness represents the line integral of the squared
density in that pixel
– Color represents the temperature
– But there is quite a bit of computation involved
Practical Results
• Program currently runs on single desktop
computer and attached storage
– 4 GB RAM, GeForce 8800 GTS, 2x750 GB disk
• Smoothly interacts with and renders 1 TB
dataset (150 million points x 128 timesteps)
• Rarely loads or renders to full depth
– Could have arbitrarily large underlying data
Future Possibilities
• Storing and accessing data via databases
– Could even do some processing in between
• Distributed rendering
• Remote rendering
• Other datasets and data types
– Meshes, volume data, medical imaging