New Release of the ELPA Eigensolver Library
A new release of the publicly available ELPA library of direct eigensolvers for symmetric (hermitian) matrices is out.
Background: The ELPA project started in 2008 with the aim to support the then upcoming Petaflop HPC systems (hence the name “Eigenvalue Solvers for Petaflop Applications”). During the last decade the ELPA-library has proven to provide highly efficient eigenvalue solvers for Petaflop systems. Nowadays, with (pre)-Exascale systems on the horizon, allthough keeping its well established name, the ELPA-library targets the Exascale range. The development of the ELPA-library has been supported by BMBF grants 01IH08007 (Dec. 2008 - Nov. 2011) and 01IH15001 (Feb. 2016 - Jan. 2019). Currently, the ELPA libray serves as an important building block by the European Center of Excellence project Nomad CoE and has been used by many HPC software packages, mainly for electronic structure calculations.
The latest release, 2023.11.001 has a particular focus on GPU performance for all major HPC platforms. It is freely available for download and use (under a LGPL v3 license) and is installed on MPCDF HPC systems. Highlights of ELPA 2023.11.001 include:
- enable gpu-streams per default for NVIDIA and AMD GPUs
- update / improved documentation and man pages
- fix compilation error on AMD GPUs
- fix SVE 256 compute kernels for ARM CPUs
- allow (currently in parts of ELPA) to use NVIDIA NCCL for device to device communication
- speed up of GPU version of hermitian_multiply by up to an factor of 4
- significantly faster full-to-tridiagonal step in ELPA 1stage GPU
- significantly faster ELPA 2stage solver on Intel GPUs
- consistent enabling/disabling of SKEW_SYMMETRIC in header files
- new setup_gpu API function