Seminars & Events
Computation Institute Lunch Presentation
"Ultrascale Visualization: A Case Study of Parallel Volume Rendering on Blue Gene/P"
DATE: November 6, 2008
TIME: 12:00 pm
SPEAKER: Thomas Peterka, Mathematics and Computer Science Division, Argonne National Laboratory
LOCATION: Room A134, Bldg. 221, Argonne National Laboratory, RI405, 5640 S. Ellis Ave, University of Chicago
Description:
Supercomputer size, computing power, and computational results are rapidly approaching petascale proportions. For example, with the recent deployment of the Blue Gene/P at Argonne’s Leadership Computing Facility, scientists can compute datasets hundreds of terabytes in size at hundreds of teraflops per second. Terabytes and teraflops will soon grow to petabytes and petaflops. Studying how simulation, analysis, and visualization can be collocated on the same leadership system has the potential benefits of eliminating data transfers and lays the groundwork for performing in situ visualization.
In this research, we examine how a popular scientific visualization technique, volume rendering, can be massively parallelized on BG/P and evaluate its performance when scaling to tens of thousands of processor cores. Volume rendering is the process of classifying data values according to color and opacity and then accumulating color and opacity along rays through the dataset, in our case a time-varying 3D scalar field from a supernova simulation. To parallelize this algorithm, we subdivide the dataset into regions and assign each region to a different processor. Each processor reads its portion of the dataset from storage and casts rays through it. Then, these partial results are composited together via a mass-exchange of completed subimages to form a final completed image.
Timing results show that these large visualizations are primarily I/O bound. Secondarily, compositing plays an increasing role as the number of processes grows. By measuring the performance of various stages of the algorithm at large system scales, we can identifying and improve these bottlenecks. These visualization experiments are the largest ever performed for this type of application, and at tens of thousands of processor cores, the distribution of I/O, rendering, and compositing time argues that systems such as BG/P can have potential benefits not only for computation but for analysis and visualization as well. This novel approach to visualization can eliminate terabytes of data movement between systems, and sets the stage for future in situ visualization, where these tasks occur during the same time as a running simulation.
Tom Peterka joined Argonne National Laboratory in 2007 after earning his Ph.D. in computer science from the University of Illinois at Chicago. His graduate research at the Electronic Visualization Laboratory at UIC in virtual reality produced a novel autostereoscopic display system for visualization. He currently researches visualization methods that scale with the increasing size and complexity of scientific data, and virtual environments to help scientists interactively engage and understand their results.
More Information:
Lunch will be provided at both locations
Upcoming talk:
November 13 - Kamil Iskra, MCS Division, Argonne National Laboratory
Save the event to your calendar [schedule.ics]
