New GPI-2 Release Improves Communication in Large HPC Systems
With the new GPI-2 Release 1.8.0, Fraunhofer ITWM is presenting a further development of its programming model for efficient and scalable HPC applications at SC26. GPI-2 enables communication and computations to be executed largely in parallel, thereby allowing even large distributed simulations to be processed efficiently. The new version supports additional network technologies and improves the launch of large computing jobs with many processes.
Another new feature is a runtime analysis function: applications can be analyzed during execution without having to recompile the program code first. The results can be evaluated directly using common visualization tools. This makes it easier to identify performance bottlenecks in large HPC applications and to optimize the programs in a targeted manner.
Gaspils Speeds up Time-Dependent Simulations
The »GaspiLS« solver library, developed at Fraunhofer ITWM, has also been further refined. In time-dependent simulations, similar systems of equations must be solved over many consecutive computational steps, for example in flow or reservoir simulations. GaspiLS specifically exploits this similarity and can reuse previously calculated structures in subsequent computational steps. This reduces the computational effort and, in particular, significantly shortens the time-consuming preparation of the individual solution steps.
A benchmark from a reservoir simulation involving approximately one million cells demonstrates the magnitude of this effect: Across the tested process numbers, GaspiLS achieved speeds up to twice as fast as the comparison methods used. For complex transient simulations, this opens up the possibility of using computational resources more efficiently and delivering results more quickly.
Fraunhofer Institute for Industrial Mathematics ITWM
