220 likes | 308 Vues
Explore interactive simulated vascular reconstruction, performance optimization, communication in distributed systems, and results & conclusions with a focus on fluid flow simulation.
E N D
High performance distributed simulation for interactive simulated vascular reconstruction Robert G. Belleman and Roman Shulakov Section Computational Science, University of Amsterdam,Kruislaan 403, 1098 SJ Amsterdam, the Netherlands.Email:(robbel|rshulako)@science.uva.nl
Overview • Interactive distributed simulation • Simulated vascular reconstruction in a virtual environment • Performance issues • Communication between distributed components • Results • Conclusions
Interactive distributed simulation • Human in the loop experimentation • interactive exploration of data/parameter spaces
Performance issues • Asynchronous, pipelined configuration • Distributed to exploit specialised resources • To increase performance: • decrease component execution times • shorten delays
Vascular reconstruction • “Traditional” treatment
Simulated vascular reconstruction • Simulated treatment planning
Simulated vascular reconstruction • Components: • blood flow simulation • visualization in a virtual environment • interaction with simulation and visualization (treatment planning) • Requirements: • Interactive system; fast response, fast update rate
Fluid flow simulation • Lattice Boltzmann Method (LBM) • Lattice based particle method • Regular lattice, similar to CT or MRI datasets • Allows irregular 3D geometry • Allows changes at run-time • Velocity, pressure and shearstress calculated fromparticle densities • Non-compressiblehomogeneous fluid,laminar flow • Spatial and temporal locality • Ideal for parallelimplementation
Fluid flow simulation • Lattice Boltzmann Method (LBM) • Lattice based particle method • Regular lattice, similar to CT or MRI datasets • Allows irregular 3D geometry • Allows changes at run-time • Velocity, pressure and shearstress calculated fromparticle densities • Non-compressiblehomogeneous fluid,laminar flow • Spatial and temporal locality • Ideal for parallelimplementation
Parallel fluid flow simulation • Performance indication:
Communication delay • 10-100Mb of data per iteration • Velocity, pressure, shear stress • May take seconds to transfer • Increasing communication throughput • Latency hiding • Decrease latency: faster response time, increased update rate • Payload reduction • Less data: shorter transfer times
Latency hiding • Multiple connections • Waiting for acknowledgements is hidden • Packets can travel through different routes • Uses CAVERN
Payload reduction • Data encoding • Decreases level of detail • Accuracy determined by type of representation • Lossy compression • Must be used with care! • Induces latency • “Standard” compression libraries (zlib) • Reduction to 10% not uncommon • Lossless compression • Induces latency
Communication pipeline • Cascade of encoding, compression and multiple socket connections
Results • Latency hiding:
Results • Communication throughput:
Conclusions • Increased throughput • In some cases over 5 times bandwidth • Time/location(/device?) independent interactive simulation • Access to high performancecomputing resources fromlow bandwidth connections • Use anywhere, at any time…onanything?
Future work • Automatic tuning of communication stages • Intelligent interaction and presentation • Grid enabling: • Provide access to distributed resources • algorithms, scanners, databases • CAVERN can be used on Globus
Thanks! • UvA, Section Computational Science • Sloot, Hoekstra, Zhao, Artoli, Merks, Shamonin • Leiden University Medical Center (LUMC) • LKEB (Reiber, vd Geest, Schaap) • Stanford University • Biomedical department (Zarins, Taylor) • SARA Computing and Networking Services See also in ICCS 2002 proceedings: • Lattice Boltzmann flow simulation: Artoli, Hoekstra (UvA/SCS) • Segmentation of MRA images: Schaap (LUMC) • Agent based solutions to interactive systems: Zhao (UvA/SCS)
Treatment planning • Interactive planning in VR
Treatment planning • Generate grids from analytical models