HENP Support
NERSC's support to the high energy and nuclear physics community includes consulting, software development, and operation of a specialized data-intensive computing facility, the PDSF (Parallel Distributed Systems Facility). The PDSF is a networked distributed computing environment--a cluster of clusters--used by six large-scale HENP investigations for detector simulation, data analysis, and software development.
The PDSF underwent a major hardware upgrade this year. We retired 32 Hewlett-Packard HP 735/9000 workstations and 24 Sun Sparc 10 workstations. The new configuration consists of four clusters--with a total of 68 Pentium II processors and a quad-CPU Sun E450 fileserver--plus eight disk vaults, each containing 66 GB of shared storage. The PDSF hardware supports multiple operating systems--Linux, Solaris, and Windows 95, 98, and NT--for maximum flexibility. We are integrating the PDSF with high-bandwidth networking, disk cache, and the HPSS storage system as an advanced prototype of the architecture necessary for future HENP data analysis.
A milestone in physics software development was reached when a NERSC-led collaboration with Brookhaven National Laboratory and the Pittsburgh Supercomputing Center ported the CERNLIB collection of codes to a parallel architecture--a large and complex task that had been attempted several times before but never completed. Born at CERN (the European Organization for Nuclear Research), CERNLIB has been developed and supplemented by HENP researchers around the world. The codes for mathematics, physics simulations, statistical analysis, user interfaces, graphics, memory management, data I/O, etc., are centrally maintained and codified by CERN in Geneva, Switzerland, and are used by probably every university physics department and HENP research center.
Principal collaborators in porting the CERNLIB physics library to the T3E were Craig Tull and Mark Durst of NERSC (shown here), as well as Jeff Porter of the Berkeley Lab Nuclear Science Division and Pavel Nevski from Brookhaven. Enabling the codes to run on parallel machines makes it possible to analyze massive datasets from the next generation of experiments.
While simulation codes have already been running on parallel machines, porting of CERNLIB to NERSC's T3E paves the way for analysis of actual data from facilities such as Brookhaven's STAR detector, which is scheduled to start data acquisition in late 1999. Finding the scientific significance in the huge quantities of data from these experiments will require the power of massively parallel processing.