2AHM September 2005
RAVE:Resource-Aware
Visualization Environment
Dr. Ian J. GrimsteadProf. Nick J. Avis
Prof. David W. Walker
Cardiff School of Computer Science
Cardiff, Wales, UK
3AHM September 2005
Presentation Structure
● Data Visualization: Pros and Cons● A Solution: The RAVE project● Demonstration of RAVE● How RAVE works● Latest results● Conclusion
4AHM September 2005
Data Visualization:Simulations
● Test theories without physically building
● Cheaper to construct new tests● Can run overnight without human intervention
● Simulations produce lots of information● But - hard to understand...
Flow ratio Area Segment23.2 #1 213.2 #34 4
... ... ...
Too much info...
Flow ratio
Sample ASample B
...or too little
5AHM September 2005
Data Visualization:Comprehension
● Solution–graphical visualization of data● View a model of the data, not the data
● Massachusetts Bay● Colours, contours,...● Easier to
comprehend● Data is now
interactiveImage courtesy of IBM ResearchGenerated with IBM Open Visualization Data Explorer
6AHM September 2005
Data Visualization:Machine Dependence
● System is often single platform● Microsoft vs. UNIX vs. Apple Mac vs. ...● Handheld vs. workstation vs. ...● Need to buy more copies of the system!
7AHM September 2005
Data Visualization:Multiple Users
● Hard to collaborate with other users● Usually – must all crowd around one machine
● Unless a large display is available● One person “driving” – others are passive● System is not assisting with collaboration
8AHM September 2005
Data Visualization:Specialist Equipment
● May require specialist computer● Capable of displaying complex data● Prohibitively expensive to own● User may need to move to machine
● Problem if only one machine● Overloaded – too slow to be usable● All displays are in use● What if it breaks?
9AHM September 2005
Data Visualization:Summary
● Pros:● Can comprehend much more information● Data is now interactive
● Cons:● Restricted to specific machine/platform● May require specialist computer● Hard for users to collaborate
10AHM September 2005
A Solution:The RAVE Project
● RAVE supports:● Various types of machine/display
● Immersadesk → workstation → PDA● Multiple machines/resources
● Resource-aware: network, machine load● Multiple users
● Resource sharing● Collaboration
● RAVE is now demonstrated...
11AHM September 2005
Demonstration
● Recorded demo● Resources:
● Windows laptop (active clients, Java)● Remote Linux/Solaris/IRIX servers
● Data servers● Uses:
● WeSC UDDI server● WeSC Service-Orientated Grid
13AHM September 2005
The RAVE Project:How it Works
● Each RAVE component now examined:● Data Distribution - Data Server● Displaying the Data - Active Client● Lightweight clients - Render Server, Thin Client● Service Discovery● Tiled rendering with Active Client● Remote (dynamic) data feed
14AHM September 2005
Data Distribution● First component: Data Server● Acts as a distribution point & interpreter
● Understands many types of data● Uses Java3D+Xj3D as importer
Data to be visualised
DataServer
Internetor remote machine
VisualizationData
RAVEClient
RAVEClient
RAVEClient
15AHM September 2005
Displaying the Data● Second component: Active RAVE Client
● “Active” – facilities to draw on its own● Accepts feed from Data Server● Presents images of data to user
VisualizationData
DataServer
Active RAVE Client
Visual drawn on local machine
Isosurface of MRI from Large Geometric Models Archive (~850kpoly, 3
nodes, 19.8Mb raw data)Bootstrap DS→AC: 12.4s
Note: Windows XPDiffusion Tensor Imaging,
SHEFC Brain Imaging Research Centre for
Scotland, Martin Connell and Mark Bastin
(~950kpoly, 2200 nodes, 29.8Mb raw data)
Bootstrap DS→AC: 20.9s
Geology dataset (10 minute ETOPO from
National Geophysical Data Center (~4.6Mpoly, 3
nodes, 109.6Mb raw data)Bootstrap DS→AC: 48.3s
16AHM September 2005
● Third component: the Render Server● Drawn visual sent to Thin RAVE Clients
● “Thin”-insufficient power/resources to draw data
Interaction
Visual
Lightweight Clients
DataServer
Thin Client
VisualizationData
RenderServer
Visual drawnoff-screen (hidden)
Isosurface of MRI scan Large Geometric Models Archive (~850kpoly, 3
nodes, 3.2fps @ 400x400 11Mbit wireless)
MolScript VRML of 1PRC molecule (Research
Collaboratory for Structural Bioinformatics –
Protein Data Bank)(~546kpoly, 29,000
nodes, 23.2Mb raw data)96.5s DS→RS (# nodes)
3.2fps @ 400x400 (11Mbit shared wireless)
17AHM September 2005
Service Discovery
● Servers are “advertised” on the network● Using standardised methods
● UDDI, Grid/Web Services● We can reuse the work of other people
● UDDI4J, Apache Axis, Globus● Human user can see list of servers
● Select most appropriate one ● Consider speed, memory, bandwidth...
● May already have your required data on it● Or automatically select with a heuristic
18AHM September 2005
Remote, Dynamic Data
● Independent simulation can supply Data Server
● Simulation code instrumented● Transmits scene creation to Data
Server● Subsequent updates also sent ● Data Server reflects updates● Multiple clients can view live
simulation
19AHM September 2005
Tiled Rendering
● If your machine can nearly cope:● Request assistance from a Render Service● Automatically select RS with heuristic● Locally render subset (tile) of data● Remainder rendered by Render Server
Visualization Data
DataServer
DrawnVisual
Render Server
DrawnVisual
Render Server
Active Client
UDDIServer
Available RS
Searchfor RS
20AHM September 2005
Tiled Rendering:Latest ResultsFPS of On-Screen (GeForce Go!) 420
vs Off-Screen (GeForce 5200)
1.0
10.0
100.0
1000.0
0 200 400 600 800 1,000 1,200 1,400
# 1,000 Vertices
Lo
g1
0 F
PS
100% on-screen
50% on-screen
50% off-screenGalleon
Elle
SkinHand
Frog
DTI
1PRC
Tiling advantag
e@ 600kv?Perfectly tri-
stripped
~29,000 nodes;
~2.2 v:p
~1.3 v:p
21AHM September 2005
Tiled Rendering:Discussion
● Is it worth it?● Only in specific circumstances:● When GPU fillrate is local bottleneck● T&L constant between 50% and 100%● Sufficient network bandwidth available
● Examples:● Hand dataset – perfectly tristripped
● GPU T&L not bottleneck 200% speedup● 1PRC – hardly tristripped (2.2 verts/poly)
● GPU T&L bottleneck 20% slowdown
22AHM September 2005
RAVE: Summary
● Data Server reads data and distributes● Active Client renders locally● Thin Client renders via Render Server● Active Client may request assistance● All resources shared where possible● Uses Java to support (most) platforms
23AHM September 2005
Conclusion
● Visualization – great!● But requires specialist hardware or software● Often not designed for multiple users
● Solution - “RAVE”● Utilise any available machines/resources● Collaborative – work from your desk
● Further information:● http://www.wesc.ac.uk/projectsite/rave/
24AHM September 2005
Acknowledgements
● Project funding: UK DTI & SGI● Diffuse Tensor Imaging dataset:
● Martin Connell and Mark Bastin, SHEFC Brain Imaging Research Centre for Scotland
● Molecule geometry:● Research Collaboratory for Structural
Bioinformatics Protein Data Bank, using MolScript● Skeletal hand:
● Large Geometric Models Archive, Georgia Institute of Technology
● ETOPO dataset:● National Geophysical Data Center (NGDC)