MAPS
Mathematical Applications with a Parallel-Beowulf System
P.I.:
James
Gentle
Co.Is:
Estela
Blaisten-Barojas
Tarek
El-Ghazawi
Rainald Löhner
John
Wallin
Edward
Wegman
Daniel B.
Carr
Stauts:
Funded by NSF,
Official Starting Day is October 1, 1999
A. Project Summary
The Institute for Computational Sciences and Informatics (CSI) at George Mason University has acquired a Beowulf-class parallel computing cluster with 134 CPUs or 67 dual-processor Pentium III 600 MHz nodes running the Linux operating system. Each node is a Dell Precision410 Workstation with 512MB of RAM and 13GB of disk. The cluster is interconnected via a Foundary FastIron fast eathernet 72 port switch. The system runs the Message Passing Interface (MPI) and the Parallel Virtual Machine systems (PVM) with C, C++, and Fortran bindings, and OpenMP for the node level parallelization if desired. Other parallel computing tools will include both High Performance Fortran (HPF) and Fortran 90 (F90). The equipment will be used to support several research projects with high computational demands including hydrodynamic modeling of interacting galaxies, gene network analysis, computational hemodynamics, analysis of fractality, self similarity, and complexity in aerogels, and new uses of immersive environments in statistical data analysis. The researchers on this proposal bring expertise in the mathematical and statistical sciences, domain sciences, and parallel computing and systems. They are actively teaching courses and advising a large number of students in computational sciences and statistics.
The Beowulf-class
parallel computing system will meet the computational requirements for these
research projects in a very cost-effective manner by the use of commodity
processors and open-source or off-the-shelf software. The total cost of the
equipment, including software and hardware is $200,000 for the 3 year duration
of the project, of which George Mason University will supply a total of $100,000
and NSF contributes $100,000. The
science investigations conducted by principal investigators are summarized
below. The system architecture,
construction, benchmarking and operation are overseen by Tarek El-Ghazawi.
b) Abstracts of Research Projects
Fractality, Self Similarity, and Complexity in Aerogels
Estela Blaisten-Barojas
The aim of this proposal is to develop cellular automata that characterize the aggregation of binary aerogels in the search for optimum conducting aerogels. These cellular automata will be of immediate help to experiments developing conducting aerogels currently underway. Fractal characteristics, dynamical exponents, structural behavior are some of the properties aimed in this study of a complex system such as the heterogeneous aerogels. Results from this proposal will shed light on the atypical electron and thermal conduction mechanisms in binary aerogels with enormous interfacial area and mesoporosity characteristics.
Computational Hemodynamics (Blood Flows)
Rainald Löhner
The flow of blood through the human body is a complicated process. The computer-assisted solution of the partial differential equations (PDEs) governing the flow of blood offers a way of understanding the precess. This approach sits between experiments and analysis: it is a "virtual experiment", capturing all the geometrical and physical detail of the real experiment, but solving the PDEs via numerical techniques on computers. The output of this "virtual experiment" delivers an almost unimaginable amount of information that includes all those quantities that are difficult to measure in real experiments, such as local velocities, shear stresses, residence times, etc. At present, a typical simulation takes 5-6 hours on a 16-processor SGI Origin 2000 machine. We hope to reduce CPU times significantly by improving algorithmic performance via preconditioning and multigrid techniques by a factor of 10 and by using the increased number of processors in the Beowulf-class machine.
Graphics Tools for Gene Network Analysis
George S. Michaels and Daniel B. Carr
Multiple classification and clustering analysis of a large number of genes are very computing-intensive tasks. Furthermore, the results of such analyses are difficult to understand when the number of genes can be as large as 10,000. Better types of graphical displays for exploring the clusters are necessary. The sprecific aims of this component of the proposal are: 1) Define a set of statistical clustering methods to identify parallel gene expression time-series data. 2) Develop scalable graphical layout characteristics for representing the relationships among all the genes from a eucaryotic genome. 3) Implement these algorithims on a scalable parallel Beowolf class machine.
Hydrodynamic Modeling of Interacting Galaxies
John Wallin
Most of our understanding of the gas dynamics in interacting galaxies (and to a lesser extent, in cosmology) relies on simulations which use Smoothed Particle Hydrodynamics (SPH). We have shown that relatively small differences in the artificial viscosity in SPH simulations create significant differences in the galaxy's morphological evolution and star formation rates. To address these problems, we plan to take advantage of recent advances in Element Free Galerkin methods which have been developed to simulate high speed collisions of solids. These methods will provide critical insight into the reliability of SPH simulations in astronomy. The simulations are computationally intensive. Based on benchmarks with systems similar the one we propose, we estimate a typical simulation of an interacting galaxy with (1 million self-gravitating particles, 128k gas particles) will take approximately 500 node hours to complete (approximately 8 hours of machine time).
Statistical Analysis in a Projection-Based Virtual Reality Environment
Edward J. Wegman and James E. Gentle
The CAVE
environment is a projection-based virtual reality environment. The computer
hardware for the CAVE are normally high-end Silicon Graphics systems with
expensive CRT-based projection systems. Earlier work on a MiniCAVE by one of the
proposers (Wegman) has focused on downscaling both the computing hardware and
the projection systems to relatively inexpensive and relative robust equipment.
Our use of the Beowulf-class computer system now will separate the tight
integration of the computers and projection equipment characteristic of CAVE and
MiniCAVE systems. The computer and the projectors will be in different buildings
and will be linked with a high-speed fiber-optic cable (already in place). Our
purpose in developing this system is to provide an effective environment for
exploratory data analysis.
