HR: 0800h
AN: IN11B-0469 [Abstracts]
TI: A System for Scripted Data Analysis at Remote Data Centers
AU: * Wang, D L
EM: wangd@uci.edu
AF: Department of Electrical Engineering and Computer Science
University of California, Irvine, Engineering Tower, Irvine, CA 92697,
AU: Zender, C S
EM: zender@uci.edu
AF: Department of Earth System Science
University of California, Irvine, Croul Hall, Irvine, CA 92697,
AU: Jenks, S F
EM: stephen.jenks@uci.edu
AF: Department of Electrical Engineering and Computer Science
University of California, Irvine, Engineering Tower, Irvine, CA 92697,
AB:
Terascale data reduction and analysis still remain elusive for most
geoscientists even as large scale geophysical models and observing
systems begin to produce petascale datasets.
Massive amounts of geophysical netCDF data remain underutilized
due to scientists' limited bandwidth and computational capacity.
Compounding the problem are the
unique characteristics of geoscientists' data analyses which are
poorly matched to traditional grid technologies.
Our system, the Script Workflow Analysis for MultiProcessing (SWAMP)
allows scientists to efficiently leverage grid resources for parallel
computation through a rich scripting syntax compatible with most
existing analysis scripts.
Although shell script languages are impossible to optimize for all cases,
data analysis scripting is well-defined sufficiently to allow inference
of dependencies and use of advanced compilation techniques.
SWAMP compiles scripts to detect potential parallelism and to dynamically
schedule for parallel execution on resources allocated by an
industry-standard grid scheduling engine.
By its unique focus on I/O and bandwidth considerations rather than
raw computational power, our system more efficiently handles the
characteristic I/O-boundedness of geoscience workflows. Although
scientists' scripts exhibit significant available parallelism, such
parallelism is poorly exploited by conventional means that manage I/O
and data locality as afterthoughts, if at all. In contrast, SWAMP
eliminates unnecessary data movement to achieve its parallel
efficiency in two ways: by providing a computational interface
integrated with data service, and by scheduling computation to
avoid I/O bottlenecks in disks and networks. This efficiency and
performance is designed to integrate with scientists' analysis
scripts-- many existing scripts require no modification to work with
SWAMP.
We test our system in uniprocessor, multiprocessor, and clustered
environments, with both simple and terascale test cases. Benchmarks
and other workflow statistics quantify not only the real performance
benefits but also the reasons that render traditional methods
unsuitable. We show that SWAMP typically modestly increases the data
source's computational load while significantly reducing bandwidth in
nearly all cases, making it suitable for installation in small lab
groups as well as large data centers. Local computation and bandwidth
requirements are thus drastically reduced, freeing the scientist to
perform exploratory analysis and discovery in wider scopes and finer
resolutions, and reducing time to discovery.
UR: http://code.google.com/p/swamp/
DE: 0520 Data analysis: algorithms and implementation
DE: 0525 Data management
SC: Earth and Space Science Informatics [IN]
MN: 2007 Fall Meeting