Vol. 17, No. 11	Nov. 13, 1997

Researchers receive grant for 'supercomputer'

This is a major victory for computer science and computer engineering," said Lori Pollock, computer and information sciences (CIS), when asked about a recent National Science Foundation award of $633,513.

Combined with $273,400 in University support, the five-year grant will bring the power of a supercomputer to researchers within the departments of CIS and of Electrical and Computer Engineering (ECE).

"Because the competition for these grants is so fierce, winning this award is a great external endorsement of the excellence of the CIS and ECE departments," added Errol Lloyd, CIS chairperson.

The winning grant proposal, submitted by a group of 13 researchers, was among only eight applications nationwide to receive funding through the agency's Computing and Information Science Engineering (CISE) Research Infrastructure Program, Pollock said.

The funds will pay for "parallel and distributed" computer systems- capable of handling massive number-crunching tasks or solving complicated problems such as human gesture recognition, Pollock reported. Specifically, she said, the grant will support 20 high-performance workstations and high-speed networking to connect the workstations within a "cluster," as well as personnel trained to maintain such sophisticated equipment.

Each of the 20 workstations will be equipped with four high-performance microprocessors with a shared memory, explained Guang R. Gao, ECE. "Every workstation will be as powerful as four high-performance, single-processor personal workstations combined, and the entire network of such powerful workstations will give us the capacity of a parallel supercomputer," Gao said.

Bringing supercomputing performance to the CIS and ECE departments will allow the researchers-and any students working with them-to "go on to another level of investigation," said Pollock, who serves as one of four principal investigators on the project, with Lloyd, Sandra M. Carberry of CIS, and Gao.

Carberry, along with departmental colleagues Chandra Kambhamettu, Daniel Chester and Keith Decker, will begin what is believed to be the first computer-based effort in the United States to use facial expressions for recognizing a person's intentions. Using supercomputing power, the researchers are developing algorithms capable of identifying the contribution of different expressions and gestures to communication. Ultimately, she said, these new computer programs might be able to analyze many forms of communication-from voice intonation to hand gestures and sentence construction-at the same time.

Such a program might someday prove useful for helping people better interact with computers, Carberry said. In the area of personal finance, for example, a naive user might interact with a computer program in order to get investment advice. "The computer system must recognize the user's intentions in order to respond intelligently and helpfully," she noted.

In the future, computer systems that recognize human gestures also may help people with disabilities communicate more effectively, she added. "Right now, the work is still fundamental," she said. "While financial management is being investigated as a kind of test bed for these algorithms, they could be applied to a broad range of problems."

Another member of the research team, Gonzalo Arce, ECE, will use the new supercomputing resources to improve the 'herky-jerky' quality of digital video images transmitted over the Internet. Video teleconferencing "requires coding, or compressing massive amounts of information," Arce explained. "You can't do that without enormous processing power."

Existing algorithms for compressing digital video images only handle "relatively simple processing," he noted. "If we could devote more computational power to these tasks, we could improve our algorithms so that the video would have a much more natural look to it." Video teleconferencing has become increasingly attractive to global corporations, as well as military officials, who hope to analyze distant battlefield conditions in real-time.

Pollock and her departmental colleague, Gagan Agrawal, will use the new NSF-funded computers to investigate compiler optimizations that improve the performance of programs running on the workstation cluster. The goal, Pollock said, is to make the compiler analyze the program and translate it in a way that takes advantage of multiple processors. In this way, she explained, programmers don't need to worry about which processor has the data and where the computation is being performed.

Compared to a typical "supercomputer-in-a-box," workstation clusters are more difficult to optimize for efficient performance, Pollock noted, "because you pay a higher penalty to communicate between workstations connected by a network than between processors in the same box." The work should help researchers make more effective use of parallel and distributed infrastructures for computationally intensive applications.

Gao, meanwhile, will continue his studies of multi-threaded computer architectures, which process many different tasks simultaneously via different "threads" or information pathways. In the past, Gao said, testing new system designs meant scheduling time on a Cray supercomputer. "I can use this new equipment to emulate large-scale parallel systems and architectures," he said, "including the proposed petaflops system," capable of processing one million-billion commands or "floating point operations" per second.

Agrawal, Carberry, Gao and Pollock will serve on the management committee for the NSF-sponsored project. Also serving on the research team will be Darren Vengroff and Ashfaq Khokhar of ECE; and David Saunders and Adarsh Sethi of CIS.

-Ginger Pinholster