A new initiative announced by the Dept. of Energy and the National Science Foundation aims to rapidly boost the nation's scientific computing capability 100-fold. The Strategic Simulation Initiative, or SSI, would by 2003 establish a national scientific network of "terascale" computers routinely capable of executing at least one trillion operations, or "teraflops," per second. Currently, most stand-alone supercomputers operate in the "gigaflops" range, capable of billions of calculations per second.

SSI advocates say the increased computational brawn should lead to significant advances in the understanding of the complex behaviors of atoms, molecules, fluids, materials, and of biological, climate and physics-related systems, such as nuclear fusion and quark-gluon dynamics. Computer modeling of such power and precision would essentially mimic reality, allowing scientists to radically improve the quality and accuracy of simulations. Jefferson Laboratory's participation in the initiative should enable its continuous electron beam to run more efficiently, provide users more power for analyzing experimental data and aid theorists in a more complete understanding of the atomic nucleus.

"The whole idea is to create very cost-effective supercomputing," says Jefferson Lab senior scientist Chip Watson, head of the Laboratory's high-performance computing project.

Currently, the Lab's Computer Center oversees a complex of 80 high-capacity networked computers providing data reduction and analysis in support of JLab's experimental program. The Center manages data volumes far in excess of those generated by any previous nuclear physics experiments conducted at other research facilities, either in the United States or abroad. Plans call for a doubling of the Center's data-handling network within the next year.

Elsewhere, the most demanding calculations still will be performed on stand-alone, state-of-the-art supercomputers consisting of a large number of very fast, interconnected microprocessors. But other calculation-intensive tasks will be handled by clusters of "symmetric multi-processor" machines, or SMPs, also known as "cluster supercomputers". Key to the operation of the clusters --- which should eventually built from commercially available personal computers --- will be parallel software not typically available to the average personal computer user.

Most personal computers are based on serial processors and algorithms that carry out instructions in a set sequence, one after another. Parallel computation derives its speed from a software strategy that subdivides calculations into discrete subtasks and then executes them simultaneously.

"The way to get the computing power is to go parallel," explains Roy Whitney, Jefferson Laboratory's administration division associate director. "A lot of this activity is already taking place in several arenas. There are a number of high-capacity centers. We plan to have a version here of what's available nationally and with our collaborating universities be a leader in serveral areas."

Close to Home

At Jefferson Lab, cluster supercomputing will eventually entail a large increase in computing power, with the goal of addressing problems requiring massive amounts of processing time. Included among the examples of calculation-intensive issues that JLab expects to address are much faster analysis of the huge volumes of data being produced by experimenters in the Lab's three experimental halls; the intensity and focus of the accelerator's electron beam; studies of the complex interplay of particles and forces within an atomic nucleus; and verification of the equations describing the quark theory of matter that underlies all of the Laboratory's nuclear physics experiments.

"Now's the time to make this push into high-performance supercomputing," Watson asserts. "The experiments are starting to move from acquiring data to analyzing data. We are designing FEL upgrades. Increasing our computing power for all these applications would be very helpful."

A small cluster supercomputer can be built for as little as $200,000 - one-tenth of the $2 million cost of a comparable conventionally designed machine. But for many problems its computational power will rival that of high-end models.

Cluster supercomputers should stimulate progress in advanced materials development, like the next generation in high-temperature alloys, as well as improved magnetic photovoltaics for solar cells and atom-size components that will comprise future computers. By permitting many more and faster calculations, cluster supercomputers should allow scientists to save enormous amounts of time when performing simulations prior to experiment design.

As an initial part of its local SSI effort, Jefferson Lab will also be working with universities across the country on fine-tuning next-generation parallel software. Lab-developed computer code could be applied to other physics-related research, including advanced accelerator design, energy production and transport and matters related to the structure and behavior of materials.

"There is nothing special about the hardware. What is special is the software we put on the machines," Watson explains. "We'll be doing what amounts to next-generation software development, pushing the state of the art in cluster supercomputing. It's very cost effective."

JLab's supercomputing foray will begin with the purchase of a cluster of workstations to perform complex computations related to a physics theory known as Lattice QCD. Once up and running, this cluster supercomputer should initially attain an application speed of more than 10 billion calculations per second. Future improvements and upgrades in hardware, application code and software libraries will likely hike speeds by at least a factor of 10 and eventually 100.

maintained by webmaster@jlab.org