The SNL ATDM Mathematical Libraries project aims to develop and integrate scalable, modular, and cross-cutting software infrastructure components for ATDM and other future exascale applications that utilize, where appropriate, ATDM Core CS components: Kokkos performance portability, Sacado/ROL embedded sensitivity analysis and optimization technology, DARMA asynchronous many-tasking, and DataWarehouse. These components include:
(a) KokkosKernels (KK): performance-portable sparse/dense linear algebra and graph kernels that utilize the hierarchical memory subsystem expected in future architectures; (b) Scalable Solvers (SS): optimal linear solver algorithms that exploit fine-grain parallelism for vector/SIMT and thread scaling and leverage next-generation execution and communication capabilities; (c) Agile Components (AC): tools for interface abstractions, discretization, time integration, and solution of nonlinear PDEs; and (d) Embedded Analysis (EA): tools for enabling advanced analysis workflows, focusing on embedded sensitivity analysis and optimization with use of derivatives for uncertainty quantification.
The project combines algorithmic R&D with delivery of interoperable software components that are expected to be crucial capabilities that will enable Sandia’s ATDM application codes to be performance portable across next-generation computing architectures such as GPUs and Xeon Phis. This work will include integration of these components into the application codes and improving their design and interfaces for mission relevant use cases.