ChAS(E)ing Hermitian dense eigenproblems with subspace iteration on large scale hybrid platforms with application to DFT

As modern massively parallel clusters are getting larger with beefier compute nodes, traditional parallel eigensolvers, such as direct solvers, struggle keeping the pace with the hardware evolution and being able to scale efficiently due to additional layers of communication and synchronization. This difficulty is especially important when porting traditional libraries to heterogeneous computing architectures equipped with accelerators, such as Graphics Processing Unit (GPU). Recently, there have been significant scientific contributions to the development of filter-based subspace eigensolver to compute partial eigenspectrum. The simpler structure of these type of algorithms makes for them easier to avoid the communication and synchronization bottlenecks typical of direct solvers. The Chebyshev Accelerated Subspace Eigensolver (ChASE) is a modern subspace eigensolver to compute partial extremal eigenpairs of large-scale Hermitian eigenproblems with the acceleration of a filter based on Chebyshev polynomials.In this talk, we report on the latest versions of the ChASE library by describing (i) its support for distributed hybrid CPU-multi-GPU computing architectures, and (ii) the very recent development of partial distribution of a combination of Householder- Cholesk-QR factorization and its impact on time-to-solution and memory footprint. Benchmarks on a modern heterogeneous cluster (JURECA-DC) based on double socket AMD Epyc Rome CPU and 4 NVIDIA GPUs per node are provided. Typical application of ChASE are sequences of large Hermitian eigenproblems as they appear in LAPW methods. ChASE is also application-code ready, in the sense that comes with a simple C++/Fortran interface that allows its integration with typical electronic structure legacy codes