1 / 35

Advanced Scientific Visualization

Advanced Scientific Visualization. Paul Navrátil 28 May 2009. Scientific Visualization. “The purpose of computing is insight not numbers.” -- R. W. Hamming (1961). Visualization Allows Us to “See” the Science. Raw Data. Geometric Primitives. Pixels. 01001101011001 11001010010101

zared
Download Presentation

Advanced Scientific Visualization

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. AdvancedScientific Visualization Paul Navrátil 28 May 2009

  2. Scientific Visualization “The purpose of computing is insight not numbers.” -- R. W. Hamming (1961)

  3. Visualization Allows Us to “See” the Science Raw Data Geometric Primitives Pixels 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 Application Render

  4. But what about large, distributed data? 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010

  5. Or distributed rendering? 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010

  6. Or distributed displays? 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010

  7. Or all three? 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010 01001101011001 11001010010101 00101010100110 11101101011011 00110010111010

  8. Schedule • Advanced Scientific Visualization Topics • Large Data Visualization Challenges • Potential Solutions • Demonstrations • Lab • Remote Visualization on Spur and Ranger • Scripting Visualization using Python in VisIt or Paraview

  9. Visualization must scale with HPC Large data produced by large simulations require large visualization machines and produce large visualization results Terabytes of Data AT LEASTTerabytes ofVis GigapixelImages Resampling,Application,… Resolution to CaptureFeature Detail

  10. Visualization Scaling Challenges • Moving data to the visualization machine • Most applications built for shared memory machines, not distributed clusters • Image resolution limits in some software cannot capture feature details • Displays cannot show entire high-resolution images at their native resolution

  11. Moving Data • How much time do you have?

  12. Analyzing Data • Visualization programs only beginning to efficiently handle ultrascale data • 650 GB dataset -> 3 TB memory footprint • Allocate HPC nodes for RAM not cores • N-1 idle processors per node! • Stability across many distributed nodes • Rendering clusters typically number N <= 64 • Data must be dividable onto N cores Remember this when resampling!

  13. Imaging Data 4096 x 2160 PNG ~ 10 MB x 360 degrees ~ 3.6 GB x 30 days ~ 108 GB x 12 months ~ 1.3 TB @ 10 fps 3.6 hours @ 60 fps 36 min Image: NASA Blue Marble Project

  14. Displaying Data Dell 30” flat-panel LCD 4 Megapixel display 2560 x 1600 resolution

  15. Displaying Data Stallion – currently world’s highest-resolution tiled display 307 Megapixels 38400 x 8000 pixel resolution Dell 30” LCD

  16. Displaying Data NASA Blue Marble 0.5 km2 per pixel 3732 Mpixel(86400 x 43200) Stallion – 307 Mpixel (38400 x 8000) Dell 30” LCD – 4 Mpixel (2560 x 1600)

  17. What’s the solution?

  18. Solution by Partial Sums • Moving data – integrate vis machine into simulation machine. Move the machine to data! • Ranger + Spur: shared file system and interconnect • Analyzing data – create larger vis machines and develop more efficient vis apps • Smaller memory footprint • More stable across many distributed nodes Until then, the simulation machine is the vis machine!

  19. Solution by Partial Sums • Imaging data – focus vis effort on interesting features parallelize image creation • Feature detection to determine visualization targetsbut can miss “unknown unknowns” • Distribute image rendering across cluster • Displaying data – high resolution displays multi-resolution image navigation • Large displays need large spaces • Physical navigation of display provides better insights

  20. Old Model (No Remote Capability) Local Visualization Resource HPC System Pixels Mouse Data Archive Display Wide-Area Network Remote Site Local Site

  21. New ModelRemote Capability Pixels Large-Scale Visualization Resource HPC System Display Mouse Data Archive Wide-Area Network Remote Site Local Site

  22. Spur / Ranger topology visqueue $SCRATCH spur $WORK $HOME Vis nodesivis[1-7|big] login3.ranger normaldevelopment<etc>queues HPC nodesixxx-xxx login4.ranger Login Nodes Compute Nodes File System

  23. Parallel Visualization • Task parallelism – passing results to 1 process for rendering Timesteps Processes

  24. Parallel Visualization Pipeline parallelism Useful when processes have access to separate resources or when an operation requires many steps. Timesteps Processes

  25. Parallel Visualization Data parallelism Data set is partitioned among the processes and all processes execute same operations on the data. Scales well as long as the data and operations can be decomposed. Timesteps Processes

  26. Parallel Visualization Libraries • Chromium – http://chromium.sourceforge.net • Sits between application and native OpenGL • Intercepts OpenGL calls, distribute across cluster • Can do either sort-first or sort-last (sort-first is simpler, sort-last can be better for large data) • Last update 31 Aug 2006, no new GL goodies • IceT – http://www.cs.unm.edu/~kmorel/IceT/SAGE – http://www.evl.uic.edu/cavern/sage/CGLX – http://vis.ucsd.edu/~cglx/ • specifically for large tiled displays • Must use IceT / SAGE / CGLX API in code • Mesa – http://www.mesa3d.org/ • Software rendering library • Enables OpenGL rendering on machines without GPUs

  27. Open-Source Parallel Vis Apps • VisIt – https://wci.llnl.gov/codes/visit/ • Good scaling to hundreds of cores • Integrated job launching mechanism for rendering engines • Good documentation and user community • GUI not as polished • ParaView – http://www.paraview.org/ • Polished GUI, easier to navigate • Less stable across hundreds of cores • Official documentation must be purchased, though rich knowledge base on web (via Google)

  28. CUDA – coding for GPUs • C / C++ interface plus GPU-based extensions • Can use both for accelerating visualization operations and for general-purpose computing (GPGPU) • Special GPU libraries for math, FFT, BLAS Image: Tom Halfhill, Microprocessor Report

  29. GPU layout Image: Tom Halfhill, Microprocessor Report

  30. GPU Considerations • Parallelism – kernel should be highly SIMD • Switching kernels is expensive! • Job size – high workload per thread • amortize thread initialization and memory transfer costs • Memory footprint – task must decompose well • local store per GPU core is low (16 KB on G80) • card-local RAM is limited (~1GB on G8x) • access to system RAM is slow (treat like disk access) • GPU N-body study (in GPU Gems 3): http://www.nvidia.com/object/io_1195170003876.html

  31. Summary • Challenges at every stage of visualization when operating on large data • Partial solutions exist, though not integrated • Problem sizes continue to grow at every stage • Vis software community must keep pace with hardware innovations

  32. Obrigado!pnav@tacc.utexas.edu

More Related