Berkeley
Lab Team Wins Third Bandwidth Challenge
The Bandwidth Challenge competition at the annual SC conference
on high performance computing and networking has been a showcase
for emerging Grid technologies and applications. At SC2002,
the Berkeley Lab team achieved their third consecutive victory
with a data transfer rate five times faster than the previous
year’s top speed. The team won top honors for the Highest
Performing Application, moving data at a peak speed of 16.8
GB/s.
The team used clusters of computers at seven sites in the
United States, the Netherlands, and the Czech Republic. Entitled
“Wide Area Distributed Simulations Using Cactus, Globus,
and Visapult,” the winning application modeled gravitational
waves generated during the collision of black holes. Participating
sites were the PDSF at NERSC and clusters at the SC2002 conference
in Baltimore, Argonne National Laboratory, the National Center
for Supercomputing Applications, the Pittsburgh Supercomputing
Center, the University of Amsterdam, and the Masaryk University
in the Czech Republic. Support was provided by the Albert
Einstein Institute in Germany, the Poznan Supercomputing and
Networking Center in Poland, DOE’s Energy Sciences Network
(ESnet), Sandia National Laboratories, SysKonnect, Hewlett-Packard,
and Force10 Networks Inc.
|
|
|
|
Figure
1 The SC2001 Bandwidth Challenge
Team included staff from Berkeley Lab’s Computational
Research (CRD), Information Technologies and Services
(ITSD), and NERSC Center divisions. Shown here are Wes
Bethel (CRD), George Smith (NERSC), Alfred Early (ITSD),
John Christman (ITSD), Eli Dart (NERSC), Brent Draney
(NERSC), Mike Bennett (ITSD), David Paul (NERSC), and
John Shalf (CRD). The SC2002 team also included NERSC’s
Cary Whitney and Shane Canon (see
photo). |
|
The team ran the Visapult
volume rendering application at SC2002 to create visualizations
from the simulations being run on the participating clusters.
The OC-192 and OC-48 lines that fed the Baltimore convention
center were aggregated into three 10 Gigabit Ethernet links
to a Force10 Networks switch that fed the HP/Compaq Linux
cluster in the Berkeley Lab booth on the show floor. The team
used a cluster of Alpha-based computers loaned by Hewlett-Packard
and SysKonnect network interface cards to put together the
winning effort.
Wes Bethel, the head of Berkeley Lab’s Visualization
Group and developer of Visapult, said that improvements in
Visapult, along with the evolving networking and Grid infrastructure
of hardware, software, and middleware, helped push the team’s
data transfer to such a high rate.
The High-Performance Bandwidth Challenge encourages teams
of researchers from around the world to use, if not swamp,
the SC conference network to demonstrate applications using
huge amounts of data. The Berkeley Lab team won the first
ever Bandwidth Challenge at SC2000, moving data at an average
of 596 MB/s over 60 minutes and hitting a peak of 1.48 GB/s
over a five-second period. At the SC2001 conference, the team
(Figure 1) took the top prize by achieving a sustained network
performance level of 3.3 GB/s.
|