Hypre's Scalable Linear Solvers project aims to create scalable algorithms for solving linear equations.
Parallel processing is critical for the numerical solution of these problems, but that alone isn't sufficient. Scalable numerical algorithms are also necessary. In general, scalability refers to using additional computational resources effectively to solve increasingly larger problems. Several factors affect scalability, including the parallel computer architecture and the parallel implementation of the algorithm. However, one important issue sometimes overlooked is the algorithm itself's scalability. Scalability describes how the total computational work requirements grow with problem size and can be discussed independently of the computing platform.
Algorithms used in today's simulation codes tend to stem from unscalable technology. The work required to solve increasingly larger problems grows much faster than linearly, which isn't optimal. Applying scalable algorithms can significantly decrease simulation times by several orders of magnitude. This reduction means that what previously required a two-day run on an MPP can now be solved in 30 minutes. Additionally, by utilizing these new algorithms at scale, codes are limited only by the size of the computer's memory, enabling researchers to solve enormous problems.
Scalable algorithms enable application scientists to pose new questions and obtain answers to existing problems. Suppose simulations with high resolutions take several days to run. In that case, a scientist may have difficulty running a refined (i.e., more accurate) model. Likewise, they might narrow the scope of a parameter study due to the lengthy runtimes associated with each simulation. However, decreasing execution time with a scalable algorithm can allow scientists to perform more simulations at higher resolutions, paving the way for more accurate scientific results.
The 2.0.0 stable release of Hypre has authored various fixes and updates to bring its documentation up to date. On the other hand, the 2.2.0 beta development release comes with numerous new features. Among them, the FEI block preconditioner was renovated. Furthermore, the CGC and CGC-E coarsening algorithms were added, and several new solver types were added to AMS, with the AMS/FEI integration also seeing major improvements. Additionally, handling of the relative change test in GMRES has been improved, the weights for Struct and SStruct weighted Jacobi have been refined, and a "make check" target that does a simple code verification test has been added. Finally, truncation for interpolation based on the number of elements has been added to the ParCSR Hybrid AMG Solver.
Version 2.0.0 / 2.2.0 Beta: N/A