A Framework for Particle Advection for Very Large Data

1. A Framework for Particle Advection for Very Large Data Hank Childs, LBNL/UCDavis DavidPugmire, ORNL Christoph Garth, Kaiserslautern David Camp, LBNL/UCDavis Sean Ahern, ORNL Gunther Weber, LBNL Allen Sanderson, Univ. of Utah

2. Particle advection is a foundational visualization algorithm • Advecting particles creates integral curves

3. Particle advection is the duct tape of the visualization world Advecting particles is essential to understanding flow and other phenomena (e.g. magnetic fields)!

4. Outline • A general system for particle-advection based analysis • Efficient advection of particles

5. Outline • A general system for particle-advection based analysis • Efficient advection of particles

6. Goal • Efficient code for a variety of particle advection workloads and techniques • Cognizant of use cases from 1 particle to >>10K particles. • Need handling of every particle, every evaluation to be efficient. • Want to support many diverse flow techniques: flexibility/extensibility is key. • Fit within data flow network design (i.e. a filter)

7. Design • PICS filter: parallel integral curve system • Execution: • Instantiate particles at seed locations • Step particles to form integral curves • Analysis performed at each step • Termination criteria evaluated for each step • When all integral curves have completed, create final output

8. Design • Five major types of extensibility: • How to parallelize? • How do you evaluate velocity field? • How do you advect particles? • Initial particle locations? • How do you analyze the particle paths?

9. Inheritance hierarchy • We disliked the “matching inheritance” scheme, but this achieved all of our design goals cleanly. avtPICSFilter avtIntegralCurve Streamline Filter Your derived type of PICS filter avtStreamlineIC Your derived type of integral curve

10. #1: How to parallelize? avtICAlgorithm avtParDomIC-Algorithm (parallel over data) avtSerialIC-Algorithm (parallel over seeds) avtMasterSlave-ICAlgorithm

11. #2: Evaluating velocity field avtIVPField avtIVPVTKField avtIVPVTK- TimeVarying- Field avtIVPM3DC1 Field avtIVP-<YOUR>HigherOrder-Field IVP = initial value problem

12. #3: How do you advect particles? avtIVPSolver avtIVPDopri5 avtIVPEuler avtIVPLeapfrog avtIVP-M3DC1Integrator avtIVPAdams-Bashforth IVP = initial value problem

13. #4: Initial particle locations • avtPICSFilter::GetInitialLocations() = 0;

14. #5: How do you analyze particle path? • avtIntegralCurve::AnalyzeStep() = 0; • All AnalyzeStep will evaluate termination criteria • avtPICSFilter::CreateIntegralCurveOutput( std::vector<avtIntegralCurve*> &) = 0; • Examples: • Streamline: store location and scalars for current step in data members • Poincare: store location for current step in data members • FTLE: only store location of final step, no-op for preceding steps • NOTE: these derived types create very different types of outputs.

15. Putting it all together PICS Filter ::CreateInitialLocations() = 0; avtICAlgorithm avtIVPSolver http://www.visitusers.org/index.php?title=Pics_dev avtIVPField Vector< avtIntegral-Curve> ::AnalyzeStep() = 0; Integral curves sent to other processors with some derived types of avtICAlgorithm.

16. Outline • A general system for particle-advection based analysis • Efficient advection of particles

17. Advecting particles Decomposition of large data set into blocks on filesystem ? What is the right strategy for getting particle and data together?

18. Strategy: load blocks necessary for advection Decomposition of large data set into blocks on filesystem Go to filesystem and read block

19. Strategy: load blocks necessary for advection Decomposition of large data set into blocks on filesystem

20. “Parallelize over Particles” • Basic idea: particles are partitioned over PEs, blocks of data are loaded as needed. • Positives: • Indifferent to data size • Trivial parallelization (partition particles over processors) • Negative: • Redundant I/O (both across PEs and within a PE) is a significant problem.

21. “Parallelize over data” strategy:parallelize over blocks and pass particles PE2 PE1 PE4 PE3

22. “Parallelize over Data” • Basic idea: data is partitioned over PEs, particles are communicated as needed. • Positives: • Only load data one time • Negative: • Starvation!

23. Contracts are needed to enable these different processing techniques • Parallelize-over-seeds • One execution per block • Only limited data available at one time • Parallelize-over-data • One execution total • Entire data set available at one time

24. Both parallelization schemes have serious flaws. • Two approaches: Parallelize over particles Parallelize over data Hybrid algorithms

25. The master-slave algorithm is an example of a hybrid technique. • Uses “master-slave” model of communication • One process has unidirectional control over other processes • Algorithm adapts during runtime to avoid pitfalls of parallelize-over-data and parallelize-over-particles. • Nice property for production visualization tools. • (Implemented in VisIt) D. Pugmire, H. Childs, C. Garth, S. Ahern, G. Weber, “Scalable Computation of Streamlines on Very Large Datasets.” SC09, Portland, OR, November, 2009

26. Master P0 P1 P2 P3 Master Master Master P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 P14 P15 Slave Slave Slave Slave Slave Slave Slave Slave Slave Slave Slave Slave Master-Slave Hybrid Algorithm • DividePEsinto groups of N • Uniformly distribute seed points to each group • Master: • Monitor workload • Make decisions to optimize resource utilization • Slaves: • Respond to commands from Master • Report status when work complete

27. Master Process Pseudocode Master() { while ( ! done ) { if ( NewStatusFromAnySlave() ) { commands = DetermineMostEfficientCommand() for cmd in commands SendCommandToSlaves( cmd ) } } } What are the possible commands?

28. Commands that can be issued by master • Assign / Loaded Block • Assign / Unloaded Block • Handle OOB / Load • Handle OOB / Send • OOB = out of bounds Master Slave Slave is given a particle that is contained in a block that is already loaded

29. Commands that can be issued by master • Assign / Loaded Block • Assign / Unloaded Block • Handle OOB / Load • Handle OOB / Send • OOB = out of bounds Master Slave Slave is given a particle and loads the block

30. Load Commands that can be issued by master • Assign / Loaded Block • Assign / Unloaded Block • Handle OOB / Load • Handle OOB / Send • OOB = out of bounds Master Slave Slave is instructed to load a block. The particle advection in that block can then proceed.

31. Send to J Commands that can be issued by master • Assign / Loaded Block • Assign / Unloaded Block • Handle OOB / Load • Handle OOB / Send • OOB = out of bounds Master Slave Slave J Slave is instructed to send the particle to another slave that has loaded the block

32. Master Process Pseudocode Master() { while ( ! done ) { if ( NewStatusFromAnySlave() ) { commands = DetermineMostEfficientCommand() for cmd in commands SendCommandToSlaves( cmd ) } } }

33. Driving factors… • You don’t want PEs to sit idle • (The problem with parallelize-over-data) • You don’t want PEs to spend all their time doing I/O • (The problem with parallelize-over-particles) •  you want to do “least amount” of I/O that will prevent “excessive” idleness.

34. Heuristics • If no Slave has the block, then load that block! • If some Slave does have the block… • If that Slave is not busy, then have it process the particle. • If that Slave is busy, then wait “a while.” • If you’ve waited a “long time,” have another Slave load the block and process the particle.

35. Master-slave in action S0 0: Read S0 S1 S3 0: Read S1 1: Pass S2 S4 1: Pass 1: Read

36. Algorithm Test Cases • Core collapse supernova simulation • Magnetic confinement fusion simulation • Hydraulic flow simulation

37. Workload distribution in parallelize-over-data Starvation

38. Workload distribution in parallelize-over-particles Too much I/O

39. Workload distribution in master-slave algorithm Just right

40. Workload distribution in supernova simulation Parallelization by: Data Particles Hybrid Colored by PE doing integration

41. Astrophysics Test Case: Total time to compute 20,000 Streamlines Uniform Seeding Non-uniform Seeding Seconds Seconds Number of procs Number of procs Particles Hybrid Data

42. Astrophysics Test Case: Number of blocks loaded Uniform Seeding Non-uniform Seeding Blocks loaded Blocks loaded Number of procs Number of procs Data Hybrid Particles

43. Summary: Master-Slave Algorithm • First ever attempt at a hybrid algorithm for particle advection • Algorithm adapts during runtime to avoid pitfalls of parallelize-over-data and parallelize-over-particles. • Nice property for production visualization tools. • Implemented inside VisIt visualization and analysis package.

44. Final thoughts… • Summary: • Particle advection is important for understanding flow and efficiently parallelizing this computation is difficult. • We have developed a freely available system for doing this analysis for large data. • Documentation: • (PICS) http://www.visitusers.org/index.php?title=Pics_dev • (VisIt) http://www.llnl.gov/visit • Future work: • UI extensions, including Python • Additional analysis techniques (FTLE & more)

45. Acknowledgements • Funding: This work was supported by the Director, Office of Science, Office and Advanced Scientific Computing Research, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231 through the Scientific Discovery through Advanced Computing (SciDAC) program's Visualization and Analytics Center for Enabling Technologies (VACET). • Program Manager: Lucy Nowell • Master-Slave Algorithm: Dave Pugmire (ORNL), Hank Childs (LBNL/UCD), Christoph Garth (Kaiserslautern), Sean Ahern (ORNL), and Gunther Weber (LBNL) • PICS framework: Hank Childs (LBNL/UCD), Dave Pugmire (ORNL), Christoph Garth (Kaiserslautern), David Camp (LBNL/UCD), Allen Sanderson (Univ of Utah)

46. Thank you!! A Framework for Particle Advection for Very Large Data Hank Childs, LBNL/UCDavis DavidPugmire, ORNL Christoph Garth, Kaiserslautern David Camp, LBNL/UCDavis Sean Ahern, ORNL Gunther Weber, LBNL Allen Sanderson, Univ. of Utah