Trade.Information

Performance increase for your physics program by parallel computing and multiple cores

Sixty-four computers connected to each other by orange wires in the belly of Building 50 made the difference between immediate and postponed glory.

BaBar physicist Brian Petersen and collaborators asked the computers to work together to analyze 1.3 million events from particle collisions. Thanks to parallel computing—where tasks are intelligently shared and therefore done more quickly—Petersen got his answers back in time to present stunning new results at a conference in March. He saved 687 days worth of wall clock time.

Now that computing speed has reached its limit—processing speeds aren't any faster than a few years ago—computer manufacturers are doubling and quadrupling the number of cores, the arithmetic unit inside a processor. "The only way to get a performance increase for your physics program is to make use of parallel computing and multiple cores," said Alf Wachsmann of the High Performance Storage and Computing team, part of the Systems group in the Scientific Computing and Computing Services (SCCS) department.

Wachsmann periodically gives lectures to SLAC scientists, encouraging them to shift their thinking when writing programs to take advantage of parallel computing.

In addition to solving problems more quickly, parallel computing allows researchers to solve more complex problems, by taking advantage of the computers' combined memory.

The Kavli Institute for Particle Physics and Cosmology has a dedicated parallel computing cluster in Building 50. In addition, KIPAC's Tom Abel uses a separate SGI parallel computer to more precisely simulate the structures that populated the early universe. The Klystron Department also has a parallel computing cluster to run simulations. The Advanced Computation Department has used parallel computing to develop a code for simulating accelerator structures

In picture:
Alf Wachsmann displays the orange wires that connect sixty-four computers (not seen in this photo) to the Myrinet switch, allowing them to intelligently share problem-solving tasks.

Release link: http://today.slac.stanford.edu/feature/2007/parallel-computing.asp