A publication of the Office of Advanced Simulation & Computing, NA-114, NNSA Defense Programs
December 2007
NA-ASC-500-07—Issue
Return to this issue’s stories
Livermore Computer Scientists Receive R&D100 Award for hypre
Lawrence Livermore National Laboratory computer scientists have developed a software library called hypre that allows researchers to use supercomputers such as BlueGene/L and ASC Purple more effectively to conduct larger, more detailed simulations faster than before. This cutting-edge technology garnered a 2007 R&D100 award from the trade journal R&D Magazine for developing advances with commercial potential among the top 100 industrial inventions worldwide for 2006.
The ability to run large, detailed simulations in less time is critical to stockpile stewardship and other DOE/NNSA national security programs that use massively parallel machines with as many as tens of thousands of processors. Physical systems are described by complicated sets of mathematical equations that must be solved on computers to simulate reality. Large systems of linear equations are ubiquitous in scientific and engineering simulation codes, and solving these systems of equations is often the most time-consuming function of a code. Consequently, robust and efficient algorithms to solve systems of linear equations are in great demand.
Hypre, which stands for “high performance preconditioners,” is a software library unique in its ability to provide solution algorithms that are effective on a wide variety of problems, easily accessible using multiple user interfaces, and effectively exploit the full computational power of today’s high-performance computers.
The challenge for parallel linear solver algorithms has been scalability. An application code is scalable if it can use additional computational resources effectively. If the size of the problem and the number of processors are increased proportionally, the goal is to keep the computing time approximately the same. In practice, however, as simulations grow to be more detailed and realistic, computing time may increase dramatically even when more processors are added to solve the problem.
Hypre provides linear solver algorithms developed specifically to be scalable on large numbers of processors. As a result, some simulation times may be reduced by orders of magnitude—by as much as a factor of 30. Simulations that previously took days can now be run in hours or less.
DOE Privacy Disclaimer | Sandia Privacy Disclaimer | SAND 2007-8167 W
Developed and maintained by Sandia National Laboratories for NA.114