MIT gets powerful new supercomputer


MIT is now home to one of the world's fastest supercomputers. Pleiades, a clustered multiprocessor computing system made possible by a $1.5 million research grant from Digital Equipment Corp., will be housed at the Laboratory for Computer Science (LCS).

Pleiades, the most powerful computer system ever installed on campus at MIT, will put formidable computing resources at the fingertips of MIT researchers from many disciplines. The supercomputer will enable advances in global climate modeling, quantum physics, natural science research, and hardware and software technologies for cluster-based multiprocessor computing.

"We are especially excited to have this machine from Digital because it comes with several ingredients that give us a big head start toward cluster-based supercomputing," said Professor Arvind of the LCS's Computation Structures Group. The clustered system, powered by seven four-processor AlphaServer 4100 systems, has the power of a large mainframe that costs many times its price. With a peak performance of 26 gigaflops and 12 gigabytes of shared memory, it is among the 500 fastest computers in the world.

Other unique capabilities of the Pleiades system include the Memory Channel interconnect that enables internode communication with a delay of less than five microseconds. The cluster-based software, such as Digital's Fortran compiler and TruCluster, support effective use of the entire cluster as a unified computing system.

"Through close interaction with MIT's researchers, Digital expects to greatly improve its technical computing capabilities, which can be passed on to customers requiring cost-effective, high-performance systems," said Samuel H. Fuller, Digital's vice president and chief scientist.

Clustered systems are a new strategy for making the power and speed of parallel processing widely available. Instead of building special-purpose hardware, cluster computers will link several SMP (symmetric multi-processor) machines into a single coordinated unit.

These SMP machines are sold in large numbers in nontechnical computing markets and often use the latest hardware and software technology. Clustering them will make super-computers more economical and software less vulnerable to obsolescence. Potential research applications include designing new drugs for HIV, testing atomic bombs with "virtual plutonium" and modeling large-scale virtual reality projects like LCS's CityScape.

"MIT is an ideal playground for the Pleiades system, because it is a hotbed of high-performance computing," said Professor Arvind. "MIT has both the users of these massively parallel computers -- very sophisticated programmers who are constantly expanding the range and scope of their investigations -- as well as world-class researchers who are developing both hardware and software for high-performance supercomputing."

Among those already filling the Pleiades calendar are research projects in earth, atmospheric and planetary sciences (EAPS), physics, the Artificial Intelligence Laboratory and LCS. Models of ocean/atmosphere circulation are being developed by EAPS jointly with LCS. Professor John Marshall of EAPS said his work "is central to our understanding of the climate of the earth and how it is changing, and the impact of phenomena like El Ni���o on both climate and weather."

In the field of quantum physics, supercomputing makes it possible to predict complex properties of materials from the behavior of individual atoms. This allows researchers to investigate the nanoscale structure of silicon, the mechanical strength of metals, and how chemical reactions proceed in water.

"The cluster gives us the power to answer important questions that cannot be answered in any other way," said Assistant Professor of Physics Tom������������������s Arias.

Professor Paul Viola of the AI Lab said, "We can now investigate problems in machine learning, smart text retrieval and visual object recognition. These problems are either difficult or impossible without high-speed parallel computers because of the massive calculations involved."

The Supercomputing Technologies Group at LCS will use Pleiades to continue to refine its award-winning CilkChess program. "It will attract a wider range of researchers who want to use CILK, and will give us more information on the practical realities of high-performance computing," said Dr. Charles Leiserson, professor of computer science and engineering.

Professor Arvind's group plans to connect the Pleiades system's processors to each other and to other processors using MIT's ARCTIC network. This fast network will enable researchers to run even larger applications.

The Pleiades system will serve as a focus for collaboration between researchers at MIT and Digital, particularly at the company's Cambridge Research Laboratory and within its High Performance Technical Computing Group in Nashua, NH.

Digital is also sponsoring a two-year research fellowship in high-performance computing at LCS. Matteo Frigo, an LCS graduate student, is the first fellow. His work will focus on ways of making computation easier and faster. He is a co-creator of FFTW, the "Fastest Fourier Transform in the West," which outperforms even vendor-supplied machine-dependent code on many platforms.

"To make computers truly useful, we will need millions of times greater performance than we have today. Clustered multiprocessors like the Digital Pleiades system are already a major landmark in handling the heaviest computational needs we have now," said Michael Dertouzos, LCS director. "As the recently explosive growth in the performance of single processors slows down, such clusters will become even more important."

A version of this article appeared in MIT Tech Talk on October 29, 1997.


Topics: Computer science and technology

Back to the top