# Hypre 2.0.0 / 2.8.0 Beta

Develop scalable algorithms

Hypre's goal of the Scalable Linear Solvers project is to develop scalable algorithms and software for solving large, sparse linear systems of equations on parallel computers.

The primary software product is hypre, a library of high performance preconditioners that features parallel multigrid methods for both structured and unstructured grid problems.

The problems of interest arise in the simulation codes being developed at LLNL and elsewhere to study physical phenomena in the defense, environmental, energy, and biological sciences.

Although parallel processing is necessary for the numerical solution of these problems, alone it is not sufficient. Scalable numerical algorithms are also required. By "scalable" we generally mean the ability to use additional computational resources effectively to solve increasingly larger problems. Many factors contribute to scalability, including the architecture of the parallel computer and the parallel implementation of the algorithm. However, one important issue is often overlooked: the scalability of the algorithm itself. Here, scalability is a description of how the total computational work requirements grow with problem size, which can be discussed independent of the computing platform.

Many of the algorithms used in today's simulation codes are based on yesterday's unscalable technology. This means that the work required to solve increasingly larger problems grows much faster than linearly (the optimal rate). The use of scalable algorithms can decrease simulation times by several orders of magnitude, thus reducing a two-day run on an MPP to 30 minutes. Furthermore, the codes that use this technology are limited only by the size of the machine's memory because they are able to effectively exploit additional computer resources to solve huge problems.

Scalable algorithms enable the application scientist to both pose and answer new questions. For example, if a given simulation (with a particular resolution) takes several days to run, and a refined (i.e., more accurate) model would take much longer, the application scientist may forgo the larger, higher fidelity simulation. He or she also may be forced to narrow the scope of a parameter study because each run takes too long. By decreasing the execution time, a scalable algorithm allows the scientist to do more simulations at higher resolutions.

The primary software product is hypre, a library of high performance preconditioners that features parallel multigrid methods for both structured and unstructured grid problems.

The problems of interest arise in the simulation codes being developed at LLNL and elsewhere to study physical phenomena in the defense, environmental, energy, and biological sciences.

Although parallel processing is necessary for the numerical solution of these problems, alone it is not sufficient. Scalable numerical algorithms are also required. By "scalable" we generally mean the ability to use additional computational resources effectively to solve increasingly larger problems. Many factors contribute to scalability, including the architecture of the parallel computer and the parallel implementation of the algorithm. However, one important issue is often overlooked: the scalability of the algorithm itself. Here, scalability is a description of how the total computational work requirements grow with problem size, which can be discussed independent of the computing platform.

Many of the algorithms used in today's simulation codes are based on yesterday's unscalable technology. This means that the work required to solve increasingly larger problems grows much faster than linearly (the optimal rate). The use of scalable algorithms can decrease simulation times by several orders of magnitude, thus reducing a two-day run on an MPP to 30 minutes. Furthermore, the codes that use this technology are limited only by the size of the machine's memory because they are able to effectively exploit additional computer resources to solve huge problems.

Scalable algorithms enable the application scientist to both pose and answer new questions. For example, if a given simulation (with a particular resolution) takes several days to run, and a refined (i.e., more accurate) model would take much longer, the application scientist may forgo the larger, higher fidelity simulation. He or she also may be forced to narrow the scope of a parameter study because each run takes too long. By decreasing the execution time, a scalable algorithm allows the scientist to do more simulations at higher resolutions.

- last updated on:
- November 16th, 2011, 11:16 GMT
- price:
- FREE!
- homepage:
- www.llnl.gov
- license type:
- Free for non-commercial use
- developed by:
**Rob Falgout**- category:
- ROOT \ Science and Engineering \ Mathematics

Hypre

**In a hurry?** Add it to your *Download Basket!*

0/5

What's New in version 2.8.0 Beta

- This version adds an Auxiliary-space Divergence Solver (ADS), a redundant coarse-grid solve option to BoomerAM, and a Euclid preconditioner option to the Fortran interfaces for the ParCSR Krylov solvers.
- It extends the AMS and ADS solvers to support (arbitrary) high-order H(curl) and H(div) discretization methods.
- It updates and refines some of the examples.
- There are assorted bugfixes.

#### Add your review!

SUBMIT