Hot Topic - UD Experts & Speakers

video

*

Media Contacts:
Public Relations

*

Off the Wire:
Story Archives

*

Experts Home

*

University of Delaware unveils new Samson supercluster computer

NEWARK, DE.--The University of Delaware has taken a quantum leap into supercomputing with installation of the world's largest Advanced Micro Devices Athlon supercluster. The supercluster features 128 processors connected by a Dolphin Interconnect Solutions high performance network and designed and assembled by the California-based RackSaver corporation.

The supercluster, which has been named Samson, is expected to crack the top 200 among the world's fastest and most powerful supercomputers, making it one of the most powerful, nongovernmental computers in the region.

In addition, the supercluster structure allows Samson to be built at a 10th of the cost of a traditional supercomputer.

"Cluster computing is a relatively new idea in supercomputing, and it is taking the world by storm," according to William Matthaeus, a professor in the Bartol Research Institute in the College of Arts and Science, which will house the supercomputer. Matthaeus led a team that won National Science Foundation funding for the project.

"The new parallel cluster computer will solidify the University of Delaware's position as a leader in new technology, providing important computing power for UD researchers and students in physics, engineering and computer sciences," Provost Mel Schiavelli said. "It will provide an unparalleled learning experience for all concerned and fits well with the University's commitment to teaching, research and technology.

"The University recognizes the value of such cluster supercomputers, which provide exceptional computational power and speed, and is proud to stand at the leading edge of this emerging technology," he added.

The University received support from Advanced Micro Devices, maker of Athlon, which donated 64 of the 128 processors used in the supercluster. "We are excited that the Samson supercomputer is showcasing the AMD Athlon processor," said Pat Moorhead, vice president of marketing at AMD's computation products group. "This demonstrates another step in AMD's continuing plan to expand into the high-performance server and workstation markets. At the same time, both the academic and private sectors are evaluating AMD Athlon processor-based supercomputers for use in the highest areas of the scientific workplace."

RackSaver served as the integrator for the supercomputer, which is housed in the UD Computing Center. "The design of the Samson supercomputer using AMD Athlon processors reflects the economic and dynamic changes in supercomputing, with customers demanding higher levels of processing power with a minimum use of space," David Driggers, president and CEO of RackSaver, said. "RackSaver specializes in delivering leading edge high density systems that incorporate best of breed components and that can fit into specific space limitations. The Samson supercomputer with AMD Athlon processors fulfills both the space and performance needs of the University of Delaware."

"We are very excited to be part of this program," Keith Murphy, Dolphin Interconnect vice president of sales and marketing, said. "This system will be our largest high performance supercluster installation in the U.S. It will also certainly be our fastest.

"Our contribution is Wulfkit, the high performance interconnect hardware, and the MPI (Message Passing Interface) software combination," Murphy said. "Wulfkit gave UD the opportunity to develop a very cost effective and scalable super computing solution using off-the-shelf computer components."

"The University is grateful for the support of the National Science Foundation, AMD, Dolphin Interconnect and RackSaver in this endeavor," Schiavelli said.

The supercomputer will be benchmarked using international standards. Matthaeus said it will be the largest Athlon supercluster and will likely rank among the world's top 200 supercomputers on the standard Linpack benchmark. When fine tuning is completed, Samson will be able to complete 200 billion arithmetic operations per second, or 200 gigaflops.

"You would have to go to Aberdeen Proving Ground in Maryland or to the Pittsburgh Supercomputing Center before you would find the next computer that could compete with this," he said. "This will not be a supercomputer you can't touch, off in some government lab."

Matthaeus said the supercomputer project is the result of a $500,000 National Science Foundation major research infrastructure grant, which was sought because physics and astrophysics researchers at Bartol were in need of equipment capable of highly complex calculations and also in need of training on equipment that can carry out parallel calculations.

Because the single unit supercomputers manufactured by such industry giants as Cray and IBM were prohibitively expensive, Matthaeus said, "We realized we would have to enter into the world of parallel computing to retain our scientific standing, and the cluster approach is the cost-effective way to do this."
In the cluster approach to parallel computing, a series of commercially available processors are networked and fine-tuned to speed communications among the various processors. The computer is fed a parallel program, including the scientist's equations to be solved.

Each processor performs the same time-consuming computations on its subset of the large data set simultaneously and sends its result to the processor acting as the manager. The processors then share their information and arrive at the answer much faster than a single processor computer could do the job.

The Dolphin network is crucial in providing a high-speed communication pathway among all the processors in the supercluster, Matthaeus said.

This approach differs from traditional supercomputing, in which the processors communicate their data and results by accessing a large amount of memory stored in a separate unit shared by all of the processors. Large shared memory supercomputers are very expensive and, thus far, are not as easily extended to accommodate more processors as supercomputers created as a supercluster of processors with their own separate memories.

With the scalable array of microcomputers, or SAM, approach, the number and power of the processors, as well as the communications network, can be sized, or "scaled," to fit the problems that will be solved on the supercluster.

The parallel supercluster of PCs concept was first developed at NASA's Goddard Space Center and named Beowulf, after the mythological hero who slew the monster Grendel. A more recent advance was in the Avalon supercluster at Los Alamos.

"Parallel computing via clusters has become the cost-effective alternative to parallel computing on the large, expensive supercomputers," Lori Pollock, associate professor in the department of computer and information sciences at UD, said. "While software tools to make this kind of parallel computing programmer-friendly are still lacking, there is enough infrastructure for utilizing a cluster of PCs or workstations as a parallel machine to make it attractive. The cluster also offers easy scalability to additional processors. These clusters are particularly important to academic researchers as they bring the power of an expensive supercomputer at an affordable price."

Matthaeus said a single unit supercomputer of power equivalent to that which UD will install would cost 10 or more times the $500,000 the University spent to build Samson.

Pollock said the supercluster will have great benefit to the University, not only in terms of computing power and access but also in cross-disciplinary collaboration among physicists, computer scientists and engineers.

"For the physicists, the parallel computing cluster offers a cost-effective platform for the researchers to take their computer simulations and large-scale numerical calculations to the next level, in terms of being able to enhance their models or gain more accuracy, reliability, throughput and sample capacity, in a reasonable time," she said. "Some of their codes suffer from ridiculously long execution times when they desire more accuracy, while others require such large amounts of space for the desired data sets that they cannot be run on a single processor. For the computer scientists, the parallel versions of the scientific codes that are developed on the cluster provide a set of interesting, real world applications for experimental studies of newly developed program analysis and optimization techniques, and the cluster itself provides the means for performance analysis of various parallel computing, code optimization and scheduling techniques."

"That is one of the most exciting aspects of this project," Stuart Pittel, acting director of Bartol, said. "Not only are we providing incredible support for the Bartol research effort but also across other units in the University."

The educational component of the project is key, and Matthaeus and Pollock designed and taught a course on parallel computing during the fall 2000 semester.

"We developed a multidisciplinary course that was unique to University of Delaware in that it brought together computer science, physics, mechanical engineering and mathematics graduate students, faculty and postdocs in both the classroom setting and on parallel computing projects," Pollock said.

Contact: Neil Thomas, (302) 831-6408, nfttwo@udel.edu
Jan. 22, 2001