More than 50 percent of the world's high-energy physicists carry out their research at CERN, the European Laboratory for Particle Physics in Geneva, Switzerland. Funded by 19 European member states, it is the world's largest scientific research center. More than 8,000 scientists, engineers and computer specialists from 75 nations collaborate to do research into the fundamental building blocks of matter and the forces that hold them together.

CERN has a unique range of experimental facilities for particle physics research, including the Large Electron Positron (LEP) Collider, which, with its circumference of 27km, is currently the world's largest scientific instrument. In addition, the laboratory has installed a powerful, 64-node RISC-based Scalable POWERparallel Systems SP for access by the scientists at CERN and researchers worldwide via the Internet.

SP: computing resource with order-of-magnitude more power
"The demands on our classical mainframe were more than it could deliver. As a result, with CERN participation, major experiments had been bringing in their own UNIX-based workstations for years, to the point where the horsepower of such experiment-dedicated equipment was 10 times more than that of the mainframe itself," recalls Harry Renshall, Public Services manager at CERN's Computing and Networking Division. "So we were looking for a general computing resource an order of magnitude more in capacity to serve as a focus for the next generation of experiments.

"We wanted a reliable architecture with an operating system we knew; one we were sure could run our FORTRAN programs," Renshall continues, "along with price/performance, scalability -- and the ability to process the data in parallel mode. After surveying the commercial parallel processor market at the time, it was clear the SP satisfied our requirements."

Thousands of experimenters worldwide will have interactive access to the CERN SP to collaborate on and execute scientific analyses. From terminals, desktop PCs and workstations, researchers log on and use the system to analyze raw experimental data, exchange electronic mail and prepare their papers. The SP is a RISC-based, scalable parallel processing system running AIX -- IBM's implementation of UNIX operating computers by linking from two to hundreds of processors that work in tandem to quickly solve complex problems. CERN's 64-node SP will be used to trim the design time on some parts of CERN's new Large Hadron Collider (LHC), which will use more than 1,000 powerful superconducting magnets to force revolving protons into head-on collisions at record-high energies.

Data-driven analyses
"In the course of a given experiment, several million lines of FORTRAN code will be written, covering various stages of analysis that could last anywhere from two to five years and involve reading terabytes of data with more and more refinement at each stage," Renshall points out. "We're a very data-driven type of research. When the LHC starts working -- we hope in 2004 -- the increase in data will be about a thousand times over what we dealt with last year."

CERN's SP will be used in part to parallelize current data mining applications. Parallel processors are ideally suited for mining data because of their ability to extract information from vast amounts of data very quickly. By assimilating data from numerous sources and isolating specific patterns, scientists can conduct more precise analyses.

"We are devoting eight SP nodes to parallel work, 16 to interactive traffic, 24 nodes to serial batch work and 16 to server functions," Renshall says. "The SP is very good for supplying many different kinds of services, and you really don't want to mix those with different tuning characteristics. Interactive users demand a certain quality of response 24 hours a day, batch users have a certain turnaround expectation -- you should be able to guarantee that."

To perform server functions, three SP nodes drive four magnetic tapes each -- 4,000 200MB tapes are mounted a week here. Three act as network servers between the SP, other servers and clients in the computer center. Two hold different major high-energy physics databases. Another runs a World Wide Web service, a master bulletin board initiated at CERN. Three are in the process of being set up to run a fairly large file system of 350GB of disk, some of it under the IBM Parallel I/O File System. One runs a 12GB Andrew File System.

Finally, one SP node runs IBM's LoadLeveler(tm) software, a workload management facility for building, submitting and processing batch jobs quickly and efficiently in a distributed systems environment. Still another is being used to test IBM's ADSTARtm Distributed Storage Manager (ADSM), a backup and archive software product for distributed, multivendor workstations and LAN file servers. And the last will be used to drive new IBM 3590 High Performance Tape Subsystem drives.

Stability and scalability
"One of the constraints was to have a single system image," Renshall notes. "We could have bought 64 RS/6000s, but that would have been much harder from the management point of view, which was very important in our decision. In addition, the SP hardware and software is intrinsically stable. For the future, we need to be able to grow incrementally, as our needs increase," says Renshall, "and with the SP that's a decision we can implement in the space of a few months."

This page illustrates how one customer uses IBM products. Many factors have contributed to the results and benefits described. IBM does not guarantee comparable results. All information contained herein was provided by the featured customer and IBM Business Partners. IBM does not attest to its accuracy.