The overall goal of Pillar 1 is to extend and develop aiCMS (ab initio computational materials science) for entire code families – not just a few codes – to be able to exploit exascale for attacking new problem categories that are not feasible on today’s top supercomputers. This is achieved by isolating and optimizing the computationally dominant tasks in code-independent open-source libraries featuring high-level, code-family specific interfaces. These code families encompass essentially all codes used in the aiCMS community. Pillar 1 of NOMAD CoE will thus extend the limits of electronic-structure theory to previously inaccessible system sizes, materials complexity, accuracy and precision, devising exascale algorithms and employing co-design in software and hardware solutions to enable novel solutions to grand challenges of exceptional scientific, societal and industrial relevance.
Why do we serve so many different codes?
The clear answer is that different codes still have certain favorite application areas, and more recent codes, which involve fewer approximations, may well develop to become more important in the coming years. The ecosystems of the code families and various specific codes have proven to be extremely valuable in the past. And we do not expect this to change in the future. Thus, the NOMAD CoE will certainly continue its inclusive tradition. Therefore, many leading developers of other codes are in the second shell.
Tasks of Pillar 1
The work of Pillar 1 focuses on libraries rather than individual codes. We achieve this by providing software libraries that can be used by all four ‘code families’ to take maximum advantage of exascale supercomputers as they become accessible. For instance, ELPA-X and Libxc-X will enable exascale DFT to be integrated into any code. Likewise, we will deliver Green-X, a library comprising cubic scaling RPA and GW approaches. Moreover, with the library CC4S-X, we will demonstrate that CC methodology can be formulated and implemented in such a way that it can be easily linked to any community code providing the basic ingredients. In fact, a majority of the required mathematical expressions can be largely abstracted from the underlying code or even basis set (code family). Consequently, we will establish and publish routines that can be fed with code-specific input through rigorously defined, high-level interfaces. The performance of all libraries will be analysed and demonstrated for each code family. Parallel to these developments we offer extensive training to support the aiCMS code adaptations required to incorporate the NOMAD libraries in an optimal fashion.
Brief description of the Work Packages involved in Pillar 1
Work Package 1: Exascale DFT
This WP has two major tasks ̶ enhancing the performance and functionality of the existing libraries ELPA and Libxc, hence allowing us to profit from and exploit the computational power of exascale platforms. It provides the opensource libraries, ELPA-X and Libxc-X, as major outcomes. The computational demand mandates close collaboration with WP8 in terms of co-design based on WP1 benchmark examples.
Work Package 2: Exascale Green-Function-Based Methods
This WP will enable the electronic structure community to apply accurate beyond-DFT methodologies, based on Green functions, to problems and systems currently out of reach using state-of-the-art computers and software. The physical properties covered by these methodologies are (1) accurate total energies based on the random-phase approximation (RPA) and on Møller-Plesset perturbation theory (MP2), (2) accurate electronic structure based on the GW approach of many-body perturbation theory, and (3) temperature-dependent effects based on many-body theory of electron-phonon coupling (EPC).
Initially, this library will be interfaced with ABINIT, exciting, FHI-aims and GPAW.
Work Package 3: Exascale Coupled-cluster Methods
This WP develops advanced libraries that perform coupled cluster (CC) theory calculations for materials on exascale platforms as they become available. All CC calculations will be based on an enhanced CC4S library, named CC4S-X. CC4S-X will require input from the employed ab initio codes considering all code families. This input includes matrix elements of non-local operators such as those provided by WP1 and WP2. We note that the current efficiency and scalability of CC4S relies on its use of the CTF, which supports an efficient distribution of tensor algebraic operations on current massively parallel supercomputer platforms. This WP will involve profiling for bottleneck detection in CC4S-X and will result in proposals for the design and optimisation of specific architectural features (co-design).