ESSEX II: Equipping Sparse Solvers for Exascale II

点击翻译按钮获取中文摘要

The ESSEX-II project will use the successful concepts and software blueprints developed in ESSEX-I for sparse eigenvalue solvers to produce widely usable and scalable software solutions with high hardware efficiency for the computer architectures of the upcoming decade. All activities are organized along the traditional software layers of low-level parallel building blocks (kernels), algorithm implementations, and applications. However, the classic abstraction boundaries separating these layers are broken in ESSEX-II by strongly integrating objectives: scalability, numerical reliability, fault tolerance, and holistic performance and power engineering. Driven by Moores Law and power dissipation constraints, computer systems will become more parallel and heterogeneous even on the node level in upcoming years, further increasing overall system parallelism. MPI+X programming models can be adapted in flexible ways to the underlying hardware structure and are widely expected to be able to address the challenges of the massively multi-level parallel heterogeneous architectures of the next decade. Consequently, the parallel building blocks layer supports MPI+X, with X being a combination of node-level programming models able to fully exploit hardware heterogeneity, functional parallelism, and data parallelism. In addition, facilities for fully asynchronous checkpointing, silent data corruption detection and correction, performance assessment, performance model validation, and energy measurements will be provided. The algorithms layer will leverage the components in the building blocks layer to deliver fully heterogeneous, automatically fault-tolerant, and state-of-the-art implementations of Jacobi-Davidson eigensolvers, the Kernel Polynomial Method (KPM), and Chebyshev Time Propagation (ChebTP) that are ready to use for production on modern heterogeneous compute nodes with best performance and numerical accuracy. Chebyshev filter diagonalization (ChebFD) and a Krylov eigensolver complement these implementations, and the recent FEAST method will be investigated and further developed for improved scalability. The applications layer will deliver scalable solutions for conservative (Hermitian) and dissipative (non-Hermitian) quantum systems with strong links to optics and biology and to novel materials such as graphene and topological insulators. Extending its predecessor project, ESSEX-II adopts an additional focus on production-grade software. Although the selection of algorithms is strictly motivated by quantum physics application scenarios, the underlying research directions of algorithmic and hardware efficiency, accuracy, and resilience will radiate into many fields of computational science. Most importantly, all developments will be accompanied by an uncompromising performance engineering process that will rigorously expose any discrepancy between expected and observed resource efficiency.