Putting Accelerated NLA to the Test: A Benchmarking Study of State-of-the-Art Libraries
Presenter
DescriptionAs high-performance computing enters the exascale era, numerical linear algebra (NLA) remains a critical performance bottleneck for scientific applications in materials science, quantum chemistry, and climate modelling. These fields require NLA solvers capable of efficiently performing operations such as matrix multiplication, Hermitian eigenproblems, and various factorizations on heterogeneous, GPU-accelerated architectures. However, performance and scalability are strongly influenced by specific matrix properties, required precisions, storage formats, and the underlying hardware abstraction.
This talk presents the findings of a comprehensive survey on NLA requirements within modern simulation codes, alongside a benchmarking study of state-of-the-art libraries – ChASE, ELPA, SLEPc, and DLA-Future – conducted on the JURECA supercomputer. Our analysis evaluates core NLA kernels across the diverse problem types identified by the ExaNLA survey, focusing on how these solvers scale on distributed, accelerated systems. The study assesses and compares library performance under standardised conditions to identify optimal approaches for specific problem classes.
The results provide a quantitative roadmap for the ExaNLA experience, highlighting the strengths and limitations of the current ecosystem. These insights will help shape the future ExaNLA library, ensuring scalability on next-generation exascale machines by synthesising problem-driven solutions with the expertise of existing state-of-the-art NLA frameworks.
This talk presents the findings of a comprehensive survey on NLA requirements within modern simulation codes, alongside a benchmarking study of state-of-the-art libraries – ChASE, ELPA, SLEPc, and DLA-Future – conducted on the JURECA supercomputer. Our analysis evaluates core NLA kernels across the diverse problem types identified by the ExaNLA survey, focusing on how these solvers scale on distributed, accelerated systems. The study assesses and compares library performance under standardised conditions to identify optimal approaches for specific problem classes.
The results provide a quantitative roadmap for the ExaNLA experience, highlighting the strengths and limitations of the current ecosystem. These insights will help shape the future ExaNLA library, ensuring scalability on next-generation exascale machines by synthesising problem-driven solutions with the expertise of existing state-of-the-art NLA frameworks.
TimeMonday, June 2916:30 – 17:00 CEST
LocationBldg. 6 – 002
Session Chair
Event Type
Minisymposium
