OptIPuter Data, Visualization, and Collaboration Research

1
OptIPuterData, Visualization, and
CollaborationResearch

• Jason Leigh
• OptIPuter Co-PI
• Associate Professor
• University of Illinois at Chicago
• September 2003

2
OptIPuter Pipeline
JuxtaView Vol-a-Tile Grid Visualization Utility
OptiStore Active Storage
Scalable Resolution Displays Continuum
Scientific Data Mining
TeraVision
LambdaRAM
Data Services for LambdaGrids
Quanta
Data Source ? Correlate ? Render ? Display
3
Data?Correlation?Visualization Pipeline
Throughput per OptIPuter Cluster Node Measured at
Each Step
JuxtaView Vol-a-Tile Grid Visualization Utility
OptiStore Active Storage
Scalable Resolution Displays Continuum
Scientific Data Mining
TeraVision
LambdaRAM
Data Services for LambdaGrids
800Mb/s
500Mb/s
Quanta
800Mb/s
Data Source ? Correlate ? Render ? Display
800Mb/s
Variable
2.4G/s
2G/s

• Currently investigating each phase of the
  pipeline.
• All services have been benchmarked independently
  and can saturate the network interfaces.
• Still need more extensive instrumentation.
• Available bandwidth is still far below what is
  needed, e.g.
• 5x3 tiled display with 1600x1400 per tile can
  require 2G/s per tile at 30fps.
• Collaborative visualization multicast is a
  bandwidth intensive OptIPuter application.

4
OptIPuter Visualization

• USC/ISIs Grid Visualization Utility,
• UIC/EVLs OptiStore, Vol-a-Tile, JuxtaView,
  GeoWall

5
Overview of OptIPuterData, Visualization, and
Collaboration Research Activities
JuxtaView Vol-a-Tile Grid Visualization Utility
Scalable Resolution Displays Continuum
OptiStore Active Storage
Scientific Data Mining
TeraVision
LambdaRAM
Data Services for LambdaGrids
Quanta
Data Source ? Correlate ? Render ? Display
Shalini Venkataraman, Naveen Krishnaprasad, Luc
Renambot, Charles Zhang, Jason Leigh, Tom
DeFanti Electronic Visualization Lab (EVL),
University of Illinois at Chicago Marcus
Thiebaux, Carl Kesselman Information Systems
Institute (ISI), University of Southern California
6
EVL OptIPuter LambdaRAM

• Visualization applications prefer memory-like
  access primitives.
• Ideally Remote memory mapped I/O (mmap()) gt
  LambdaRAM.
• Need aggressive remote prefetching to overcome
  latency over long distance accesses.
• LambdaRAM needs to consider the following issues
• CPU bus bandwidth overhead incurred as a result
  of aggressive prefetching.
• Amount of network bandwidth it can use for
  prefetching without negatively impacting other
  application traffic.
• Needs to monitor application memory access
  patterns to so that it can prefetch the right
  data just in time.
• Delays in the disk system that will impact
  performance.
• Ideally there should be a hierarchy from CPU to
  local memory to remote memory to remote disk.

7
LambdaRAM Preliminary Performance Results

• Comparison between Best Case and Worst Case
  Performance
• Local cache size is 128MB
• Best Case sequential access of data gt hit
  ratio is high in local cache.
• Worst Case random access of data (access of
  remote RAM is needed- hence bottleneck is
  network).

• Able to maximize access of the network.
• Performance is better than local disk system.
• Need to examine different prefetching algorithms
  based on monitoring application memory access
  patterns.

Slower performance due to overhead of lots of
small accesses
Performance drops as access request approaches
local cache size
8
EVLs JuxtaView Viewing Extremely
High-Resolution Data on the GeoWall2

• Data sets have a real need for display
  resolution.
• JuxtaView copies data across all cluster nodes as
  memory-mapped files.
• Next phase is to use LambdaRAM for remote memory
  access.
• Need to examine JuxtaViews memory access
  patterns to provide heuristics for LambdaRAM
  prefetching.

Scripps Bathymetry and digital elevation
NCMIR microscopy (2800x4000 24 layers)
9
JuxtaView Washington DC Aerial Photograph (the
need for resolution)
USGS (OptIPuter partner) Aerial photography for
Homeland Defense 350,000x350,000-pixel images
of 350 US cities, 50TB of data (Brian Davis)
10
(No Transcript)
11
JuxtaView Ocean and Lake Core Drilling Samples
Emmi Ito- U. Minnesota, Frank Rack- Joint
Oceanographic Institutes
12
ISI Visualization over a Routed Network

• Sort-first rendering sort data based on the
  portion of screen it will occupy.
• Generally considered the only practical way today
  of doing large scale data visualization on
  scalable resolution displays.
• Remote data needs to talk to several
  visualization nodes at a time.
• Problem nodes are typically imbalanced.

Displays
Viz Node
Data Node
Viz Node
Data Node
Viz Node
Data Node
Viz Node
Routing is needed to sort data to correct viz
nodes
13
EVL OptiStore and Vol-a-Tile A Visualization
Network using Photonic Switches

• Sort-Last is an alternative to Sort-First, but
  Sort-Last is traditionally considered
  impractical due to large bandwidth requirements
  in the Compositing step.
• Photonic network infrastructure makes Sort-Last
  easy. But difficult for Sort-First.
• In Sort-Last, Load is evenly balanced across all
  nodes.

Problem latency accumulates here Solution use
hardware compositors
Data Node
Viz Node
Displays
Compositing Node
Data Node
Viz Node
Compositing Node
Viz Node
Data Node
14
Rendering to Multiple Screens means Replicating
the Computing Infrastructure
Viz Node
Displays
Composite Node
Viz Node
Composite Node
Data Node
Viz Node
Data Node
Viz Node
Composite Node
Data Node
Viz Node
Composite Node
Viz Node
Routing or photonic multicasting needed here
15
Photonic Multicast Service
Glimmerglass Photonic Multicast Extension allows
for 2 x 14 multicasting
16
Photonic Multicasting a Visualization
17
EVL OptiStore and Vol-a-Tile A Visualization
Network using Photonic Switches

• OptiStore
• Management
• loading
• meta-information
• storing
• Processing
• converting
• sampling
• cropping
• Representation
• voxels
• isosurfaces
• point clouds
• Transport
• photonic reservation
• TCP
• RBUDP
• Vol-a-Tile
• Transparent access

18
OptIPuter GeoWall2 showing Vol-a-Tile
Canonical Voxel Visualization Test on the
Geowall2
19
Vol-a-Tile 3D Seismic Reflectivity Across E.
Pacific Rise fromAnatomy of a Ride-Axis
Discontinuity experiment. 1001x801x801 32bit
SIO/IGPP - Graham Kent
20
Vol-a-Tile Node of Ranvier - 1960x2560x410 8 bit
NCMIR David Lee, Mark Ellisman
21
Vol-a-Tile Time-Varying Seismic VolumeSeismic
Wave Propagation Simulation of 1994 Bolivia
Earthquake
Orthogonal Views
Navigational Volume
Sub-Volume as voxels
Sub-Volume as isosurfaces
U. Michigan Peter van Keken
22
ISI Active Storage and Grid Visualization Utility
Visualization using an experimental point
cloud technique
Active Storage performs filtering distribution
of visual data object
23
Point-Cloud Rendering of a Cricket
X-ray tomography Argonne National Lab
24
OptIPuter Display Systems for Education and
Outreach From GeoWall to GeoWall2

• GeoWall Low cost 1 megapixel passive stereo
  display using commodity PCs for displaying 3D
  data. (Paul Morin, Jason Leigh, Peter van Keken)
• Widespread adoption by GeoWall Consortium for
  research education in the Geosciences
  (www.geowall.org) (200 in 2 years)
• GeoWall2 Scalable resolution coupling with
  OptIPuter allows GeoWall users to visualize
  larger data sets.

25
OptIPuter Data Research

• UIC LACs Photonic Data Services Stack
• UCIs Scientific Data Mining
• UIC EVLs OptiStore, LambdaRAM and Quanta Toolkit

26
OptIPuterData Research Activities
JuxtaView Vol-a-Tile Grid Visualization Utility
OptiStore Active Storage
Scalable Resolution Displays Continuum
Scientific Data Mining
TeraVision
LambdaRAM
Data Services for LambdaGrids
Quanta
Data Source ? Correlate ? Render ? Display
Bob Grossman Laboratory for Advanced Computing
(LAC) University of Illinois at Chicago
27
LAC Developed OptIPuter Data Services

• Providing a data service to allow scientists to
  publish and retrieve data in the same way the
  publish web pages.
• Integrating Northwestern EVLs photonic path
  reservation services to establish dedicated
  network paths between OptIPuter data sources.
• Provided interfaces for querying and performing
  large database accesses and joins (DSTP v3).
• Developed transport protocols for moving data
  rapidly and fairly over photonic networks
  (SABUL). Refer to Joe Bannisters talk.
• Year 1 Handling mainly large multivariate
  tables.
• Year 2 Will handle 2D 3D volumes.

28
OptIPuterData, Visualization, and Collaboration
Research Activities
JuxtaView Vol-a-Tile Grid Visualization Utility
OptiStore Active Storage
Scalable Resolution Displays Continuum
Scientific Data Mining
TeraVision
LambdaRAM
Data Services for LambdaGrids
Quanta
Data Source ? Correlate ? Render ? Display
Padhraic Smyth University of California Irvine
(UCI)
29
UCI OptIPuter Scientific Data MiningClustering
Dynamic Data Sets

• Crucial to large scale data visualization.
• Developing general theoretical framework and a
  set of algorithms for statistical modeling and
  clustering of sets of curves and trajectories.
• Methodology performs curve clustering and curve
  alignment simultaneously and optimally. Prior
  work in this area relied on separate (and
  suboptimal) steps of alignment and clustering.
• Testing on relevant geoscience applications- such
  as predicting storm trajectories.
• Just completed building a 3x3OptIPuter
  visualization cluster.
• Year 2 Will apply techniquesto NCMIR/SIO data
  sets usingthe OptIPuter.

30
Year 2 OptIPuter Data, Visualization and
Collaboration Research

• Experimentation with Photonic Multicasting for
  data and visualization replication in
  collaborative scenarios.
• Extension of visualization tools to support
  collaboration in Amplified Collaboration
  Environments.
• Expanding resolution of tiled display from 20 to
  30million pixels using higher resolution
  commodity displays.
• Expansion of photonic data services to handle 2D,
  3D 4D data volumes.
• Apply data mining algorithms to NCMIR/SIO data on
  OptIPuter nodes.

31
Glossary of Research Projects

• Data Services Distributed Data Servers,
  interdomain lambda services (via ODIN) Data
  Transport over wide area. (LAC)
• Quanta Both wide area and local area data
  distribution middleware, intradomain photonic
  signalling and multicasting (EVL)
• LambdaRAM Memory-mapped I/O abstraction with
  aggressive data prefetching of remote data
  sources. (EVL)
• OptiStore Data filtering mechanism for
  converting scientific data into visual objects.
  (EVL)
• ActiveStorage Similar to OptiStore but based
  heavily on current Grid services. (ISI)
• Grid Visualization Utility Grid Visualization
  utility. Examining point-cloud-based rendering to
  provide fast interactivity. (ISI)
• Scientific Data Mining Statistical methods for
  modeling and clustering sets of curves
  trajectories. (UCI)
• JuxtaView Collaborative visualization of high
  resolution, time-varying, digital montages across
  Scalable Resolution Displays. (EVL)
• Vol-a-Tile Collaborative visualization of high
  resolution, time-varying, 3D volume data on
  Scalable Resolution Displays and Stereoscopic
  Displays. (EVL)
• TeraVision High speed, multi-directional,
  graphics streaming and multicasting to support
  collaboration. (EVL)
• Scalable Resolution Displays Displays
  constructed using LCD panel arrays and driven by
  clusters.
• Continuum Interaction and collaboration in rich
  display environments. (EVL)

32
Questions?
www.evl.uic.edu/cavern/optiputer