Unm 021400

21
Scientific Visualization Research at the Advanced Computing Lab Patrick M c Cormick Advanced Computing Lab Los Alamos National Laboratory ([email protected])

description

xxx

Transcript of Unm 021400

  • Scientific Visualization Research at the Advanced Computing Lab

    Patrick McCormickAdvanced Computing LabLos Alamos National Laboratory([email protected])

  • OverviewHigh-performance visualization.

    Large data visualization.

    End-to-end solutions.

    Algorithmic solutions.Multi-resolution.Streaming.Feature detection.

  • High-PerformanceWhy?Interactivity is one of the most important issues for successful visualization.Lowest acceptable rate: 5 frames/sec.

    How?Hardware accelerated graphics.Commercial software packages available.Custom software.

  • High-Performance

    Simulations are always pushing to higher and higher resolutions.Grid resolutions almost always increase with the capabilities of computing platform. Leads to larger and larger data sets. Data analysis/visualization performance becomes even more critical.Very few (if any) commercial solutions.

  • Example Data Sets50km resolution.# time steps: days

    5km resolution# time steps: months to years.Must resolve eddies details,10 to 300km in size,for thermal energy transfer.North Atlantic: 1280x992x40 cellsGlobal (future): 3600x2400x40 cells

    ATMOSPHERIC MODELINGOCEAN MODELING128x128x128 cells

  • ExampleSatellite ImageSimulation Data

  • Large Data VisualizationThe ocean data sets are actually only moderately sized.512x512x512 and larger data sets are becoming common.Must leverage all hardwareresources or develop newtechniques.

  • End-to-End SolutionsOptimize code for hardware specific features.The same resources needed for computation are often also necessary for visualization.Often referred to as BRUTE FORCETrade off between advanced algorithm development time and the no so elegant approach.Requires knowledge from several areas:Systems: I/O, hardware architecture, etcParallel computing.Computer graphics.Scientific visualization.

  • End-to-End SolutionsExample ApplicationVolume render time varying, large data sets (1024x1024x1024) at 5Hz.Software volume rendering is too slow (even in parallel). Parallel hardware rendering is the only choice.Requires: 5GB/sec. of data off disk.Specific hardware configurations.Custom code.

  • Volume RenderingVisualization technique for scalar fields.Treat 3D field as a semi-transparent medium.Color and Transparency controlled by scalar field values Transfer Function.Classic Algorithms: Ray Casting, Splatting.Hardware graphics accelerators now have texture capabilities:Some have only 2D textures.Latest OpenGL standard includes 3D texture.Both can be use for volume rendering.

  • Volume Rendering3D Texture ApproachSample field data on a regular grid.Store regularly sampled field in 3D texture memory.Slice the field with many slice planes (or spheres) perpendicular to line of site.Map each slice through transfer functions for color and transparency.Composite slices together to form final image.

  • End-to-End SolutionsN time steps stored in memory.Example:

  • End-to-End SolutionsExample:(in memory)(compositors)

  • End-to-End SolutionsAdvantages:FAST! Solves the high-performance requirements.Happy users.Often the only way to achieve requirements.

    Disadvantages:Software can be tricky to develop.Can be a software engineering nightmare.In many cases it still may not be fast enough. We have to be smarter...

  • Software TechniquesSoftware components/frameworks:Standard software engineering advantages..Parallel & distributed algorithms are necessary.ACL used VTK as a starting point.

  • Algorithmic TechniquesMulti-resolution techniquesUse high-resolution (complete) data only when necessary.

    Streaming.Pull in a region at a time for processing.Part of VTK.

    Feature detection.Find the interesting features in the data.

  • Multi-resolution & Streaming Example Use a hierarchical representation (adaptive mesh). Additional levels may be streamed in for viewing when necessary. Extension to 3D using octrees.

  • Feature DetectionProblem: Find the conveyor belt in the simulated data fromthe POP and MICOM ocean models. One possible solution tothis problem is currently done manually over a period ofweeks.

  • Feature DetectionGoal to find a connected global flow is challenging: Flow can be fast or slow, wide or narrow.Data can be noisy making it difficult to detect features.Techniques from computer vision and image processing may be used to correlate flow (vector field) data.

    Using a mask we visit eachcell of the field and ask, Howmany of my neighbors areheaded in a similar direction?

    Results in a scalar field that measures flow correlation.

  • Feature DetectionCurrent results highlight well known flow features.

  • The EndHigh-performance (interactive) visualization should always be a goal.More efficient (happy users).Often complicated by data set sizes and algorithm performance.If computing resources insufficient then try multi-resolution, streaming, or feature detection (or perhaps all three).