Under development since the late 1990s, the hypre library of linear solvers has been used by research institutions and private companies to simulate groundwater flow, magnetic fusion energy plasmas in tokamaks and stellarators, blood flow through the heart, fluid flow in steam generators for nuclear power plants, pumping activity in oil reservoirs, and more.
Within the US Department of Energy’s Exascale Computing Project, a team from Lawrence Livermore National Laboratory (LLNL) is preparing hypre to take advantage of the capabilities of the upcoming exascale computing systems. Specifically, the hypre project is delivering scalable performance on massively parallel computer architectures to positively impact a variety of applications that need to solve linear systems.
Benefits of hypre
The hypre library provides advanced scalable parallel linear solvers and preconditioners with a major focus on multigrid algorithms. Issues of robustness, ease of use, flexibility, and interoperability have also been important.
An additional attractive feature of hypre is the provision of conceptual interfaces, which include a structured and a semi-structured interface in combination with the traditional linear-algebra-based interface. These allow access to efficient solvers that can take advantage of structures. They also allow users to apply hypre in the ways they naturally think about their problems.
The algebraic multigrid preconditioner in hypre is scalable to more than one million processes.
The hypre team is continuing to add new multigrid methods capable of solving more-complicated problems. The team has also ported the structured solvers and selected components of the unstructured solvers to GPUs.
A new mixed-integer option was recently added that performs similarly to the default 32-bit integer version but is capable of solving larger problems. It uses less memory and is about 20%–25% faster than the 64-bit integer version.
Providing scalable linear solvers and preconditioners for massively parallel architectures to enable efficient performance of applications that need to solve linear systems can take a large portion of a simulation. Accordingly, the hypre team is in the process of porting algorithms to heterogeneous architectures while gradually increasing functionality on such architectures and improving performance on GPUs.
Researchers (all are from LLNL)
Ulrike Meier Yang (principal investigator)