Berkeley Lab Demonstrates Leadership in High-Performance Computing and Networking at SC2000

The Berkeley Lab Computing Sciences organization again demonstrated its achievements and leadership in high-performance computing and networking at the annual SC conference held last week in Dallas.

The SC conference, formerly known as Supercomputing, drew 5,200 attendees and 153 exhibitors, including vendors and research organizations, such as the DOE labs. Berkeley Lab's booth highlighted the release of new software, the development of technologies to advance DOE's Science Grid linking computers and facilities around the country, and applications for computational science. Traffic at the booth was the highest in the Lab's participation in the past six conferences.

"This year we focused on those areas where we have recognized leadership and used computer demonstrations, informative posters and a strong lineup of technical talks in our booth to reinforce our leadership position," said NERSC Division Director Horst Simon. "Based on comments from DOE program managers, conference attendees and our own staff, this was our most successful SC to date."

Highlights of Lab participation at the conference included:

NERSC Deputy Division Director Bill Kramer led the effort to design and deploy a network with massive connectivity to support high-bandwidth applications at the conference. The conference network, called SCinet, offered a combined capacity more than 196,000 times faster than a typical residential Internet connection and 200 times as fast as the connections used by many universities. Kramer was interviewed for the regional edition of the Wall Street Journal and by local television and radio news reporters.

A Berkeley Lab team made up of Brian Tierney, Wes Bethel, Jason Lee and Dan Gunter won the "Fastest and Fattest" prize for best overall performance in the SC2000 Network Challenge for Bandwidth-Intensive Applications. This was a new competition designed to demonstrate high-bandwidth applications and push the limits of network capacity. In demonstrating Visapult, Bethel's prototype application and framework for remote visualization of terascale datasets, the team recorded a peak performance level of 1.48 gigabits per second over a five-second period. Overall, the team transferred 262 gigabits of data in 60 minutes from LBNL to the SC2000 show floor in Dallas

Another winner in the Network Challenge, taking top honors as the "Hottest Infrastructure" application, was a team representing the University of Southern California/Information Sciences Institute and Argonne and Lawrence Livermore national laboratories, with help from Berkeley Lab's Arie Shoshani and Alex Sim, running "A Data Management Infrastructure for Climate Modeling Research."

Berkeley Lab released three software packages at the conference and all 300 copies were gone within two days. The first package, Berkeley Lab AMR, or Adaptive Mesh Refinement, serves as a "numerical microscope," allowing researchers to "zoom in" on the specific regions of a problem that are most important to its solution. AMR has been under development for more than 15 years by scientists in NERSC's Center for Computational Sciences and Engineering and Applied Numerical Algorithms Group (ANAG). The Visualization Group has also been actively working with ANAG on developing a new AMR visualization tool.

The second software package, developed by NERSC's Future Technology Group, contains two pieces of software to allow very high performance over cluster networks. The first, called M-VIA, is a modular version of VIA (Virtual Interface Architecture), an industry standard, high performance communication interface for system area networks. First released in late 1998, M-VIA was the first implementation of VIA for Linux. The second piece of software on the CD is MVICH, an implementation of MPI (Message Passing Interface) for VIA.

The third software package was Akenti, an authorization system designed to address issues raised by granting users access to distributed computing and networking resources controlled by multiple stakeholders.

LBNL scientists also presented three papers at the conference. NERSC's Adrian Wong, Lenny Oliker, Bill Kramer and David Bailey presented "ESP: A System Utilization Benchmark;" The Visapult team (see above) spoke on "Using High-Speed WANs and Network Data Caches to Enable Remote and Distributed Visualization;" and Bill Johnston was a co-author of "Computing and Data Grids for Science and Engineering."

A fourth paper, "A Comparison of Three Programming Models for Adaptive Applications on the Origin 2000," by Hongzhang Shan of Princeton University and co-authored by NERSC's Lenny Oliker, was recognized as Best Student Paper at SC2000.

Lab scientists led two tutorials. Manfred Zorn, Sylvia Spengler and Inna Dubchak of NERSC's Center for Bioinformatics and Computational Genomics, led an all-day tutorial on "Computational Biology and High Performance Computing." The session looked at the necessary steps for utilizing data from the human genome project and how high-performance computing can help transform the current visionary thinking into reality. A half-day tutorial on "Performance Tuning and Analysis for Grid Applications" was taught by LBNL's Brian Tierney and Dan Gunter and focused on how to find and minimize bottlenecks in distributed computing systems. As more and more computing efforts draw on distributed resources, analyzing and improving the performance of such widely distributed systems becomes increasingly important.

Other Computing Sciences employees who served as members of the conference committee include Bob Lucas, Program Committee; Jed Donnelley, Proceedings chair; Tom DeBoni, backup Proceedings chair; and Jon Bashor, Media Relations chair.