ViSUS: Visualization Streams for Ultimate Scalability

Pascucci, V

Lawrence Livermore National Laboratory

关键词: Computer Graphics;    Monitoring;    99 General And Miscellaneous//Mathematics, Computing, And Information Science;    Computers;    Processing;   

DOI  :  10.2172/15014677
RP-ID  :  UCRL-TR-209692
RP-ID  :  W-7405-ENG-48
RP-ID  :  15014677

美国|英语

来源: UNT Digital Library

【 摘 要 】

In this project we developed a suite of progressive visualization algorithms and a data-streaming infrastructure that enable interactive exploration of scientific datasets of unprecedented size. The methodology aims to globally optimize the data flow in a pipeline of processing modules. Each module reads a multi-resolution representation of the input while producing a multi-resolution representation of the output. The use of multi-resolution representations provides the necessary flexibility to trade speed for accuracy in the visualization process. Maximum coherency and minimum delay in the data-flow is achieved by extensive use of progressive algorithms that continuously map local geometric updates of the input stream into immediate updates of the output stream. We implemented a prototype software infrastructure that demonstrated the flexibility and scalability of this approach by allowing large data visualization on single desktop computers, on PC clusters, and on heterogeneous computing resources distributed over a wide area network. When processing terabytes of scientific data, we have achieved an effective increase in visualization performance of several orders of magnitude in two major settings: (i) interactive visualization on desktop workstations of large datasets that cannot be stored locally; (ii) real-time monitoring of a large scientific simulation with negligible impact on the computing resources available. The ViSUS streaming infrastructure enabled the real-time execution and visualization of the two LLNL simulation codes (Miranda and Raptor) run at Supercomputing 2004 on Blue Gene/L at its presentation as the fastest supercomputer in the world. In addition to SC04, we have run live demonstrations at the IEEE VIS conference and at invited talks at the DOE MICS office, DOE computer graphics forum, UC Riverside, and the University of Maryland. In all cases we have shown the capability to stream and visualize interactively data stored remotely at the San Diego Supercomputing Center or monitor in real-time simulation codes executed on a cluster of PC's at LLNL.