fast isosurface visualization on a high-resolution scalable display wall

Fast Isosurface Visualization on aHigh-Resolution Scalable Display Wall

Adam Finkelstein

Allison Klein

Kai Li

Princeton University

Sponsors: DOE, Intel, NSF

Overview

The display wall environmentMotivationChallenges

Isosurfaces on the display wallExtractionRenderingFuture directions

Some snapshots

Some snapshots

Some snapshots

Some snapshots

Scalable low-cost display wall

32-node I/O cluster PCs w/3-D accelerators

...

64-CPU cluster

Commodityprojectors

128-CPUProduction cluster

ExtensibleRouter

T1PPPL

Wireless links

AT&T LabLAN vBNS

Technology trends

display resolution (< 5%/year)

Time

CPU and graphics hardware perfo

rmance

memory density, disk

density (5

0-60%/year)

Scalable low-cost display wall

Now:8’ × 18’ rear-projection screen8 polysilicon LCD projectors deliver

6 million pixels per frame (4096 x 1536)A network (Myrinet) of 15 Pentium-II 450Mhz

(8 have Intergraph graphics accelerators)

Soon:15 new-generation projectors will deliver

20 million pixels per frame (6400 x 3072)A network of new-generation PCs with

new-generation 3D graphics accelerators

Multi-projector displays

SGI-based displaysGovernment labs:

ANL, LANL, LLNL, Sandia Industry:

AT&T, Panoram Tech, Trimension, ...Universities:

Minnesota, Stanford, UI Chicago, UNC

PC-based displaysPrinceton, Intel, ANLNext: Illinois, LLNL, Sandia, Lucent, ...

First Video

Research challenges

Parallel rendering Fast communication Seamless imaging Interaction techniques Spatialized sound Virtual environments Visualization systems

Visualization of isosurfaces

Goals

Large data setsVisible womanAstrophysical simulations

Large display InexpensiveHigh resolution

Interactive ratesExtractionRendering

Runtime components

Extraction

Find voxels containing the isosurface.

Communication

Send surface information to display.

Rendering

Draw the surface.

Runtime architecture

database display

network

Extraction Communication Rendering

Extraction on one processor

Acceleration methods [Cignoni97]:

Spatial -- e.g. octree [Parker,Shen]Seed -- e.g. seed and traverse [Bajaj]Value -- e.g. interval tree [Cignoni]

Extraction on one processor

Acceleration methods [Cignoni97]:

Spatial -- e.g. octree [Shen]Seed -- e.g. seed and traverse [Bajaj]Value -- e.g. interval tree [Cignoni]

–We use filtering search [Chazelle86]

Filtering search

0.00

0.12

0.38

0.57

0.61

0.78

0.93

Benefits of filtering search

Nice space / time tradeoff

Better asymptotic worst case

Very easy to code

Trivially parallelizeable

Runtime architecture

database display

network

Extraction Communication Rendering

Runtime architecture

database display

network

Extraction Communication Rendering

Communication

Gigabit network (Myrinet)

Scalable

Virtual memory mapped communication

Currently we ship voxels:voxel IDmarching cube caseedge interpolants

Runtime architecture

database display

network

Extraction Communication Rendering

Rendering

Rely on PC graphics cards

Static screen-space partitioning

Current bottleneck

Edge blending

Second Video

How do we make it faster?

Rendering:Next generation of graphics cardsLoad balancing

General:Surface simplificationMultiresolution representations

Broader directions

Other vis techniques

Remote visualizationCompressionNetworking: PPPL, AT&T, CorridorOne

Scalable storage server3 TB storage1.5 GB / sec Intelligent caching$150K