ESSEX: Equipping Sparse Solvers for Exascale Andreas Alvermann 1 , Achim Basermann 2 , Holger Fehske 1 , Martin Galgon 3 , Georg Hager 4 , Moritz Kreutzer 4 , Lukas Kr¨ amer 3 , Bruno Lang 3 , Andreas Pieper 1 , Melven R¨ ohrig-Z¨ ollner 2 , Faisal Shahzad 4 , Jonas Thies 2 , and Gerhard Wellein 4 1 Ernst-Moritz-Arndt-Universit¨ at Greifswald, Greifswald, Germany 2 German Aerospace Center, K¨ oln, Germany 3 Bergische Universit¨ at Wuppertal, Wuppertal, Germany 4 Friedrich-Alexander-Universit¨ at Erlangen-N¨ urnberg, Erlangen, Germany Abstract. The ESSEX project investigates computational issues arising at exa- scale for large-scale sparse eigenvalue problems and develops programming con- cepts and numerical methods for their solution. The project pursues a coherent co-design of all software layers where a holistic performance engineering process guides code development across the classic boundaries of application, numerical method, and basic kernel library. Within ESSEX the numerical methods cover widely applicable solvers such as classic Krylov, Jacobi-Davidson, or the recent FEAST methods, as well as domain-specific iterative schemes relevant for the ESSEX quantum physics application. This report introduces the project structure and presents selected results which demonstrate the potential impact of ESSEX for efficient sparse solvers on highly scalable heterogeneous supercomputers. 1 Sparse Solvers for Exascale Computing Energy-efficient execution, fault tolerance (FT), and exploiting extreme levels of paral- lelism of hierarchical and heterogeneous hardware structures are widely considered to be the basic requirements for application software to run on future exascale systems. Specific hardware structures and best programming models for the exascale systems are, however, not yet accessible, let alone settled. Thus, development of exascale appli- cations can be considered as a research project on its own. Existing software structures, numerical methods, and conventional programming and optimization approaches need to be reconsidered. New techniques such as FT or parallel execution on heterogeneous hardware have to be developed. A wide range of sparse linear algebra applications from quantum physics to fluid dy- namics have already identified urgent problems which can only be solved with exascale computers. The relevant sparse linear solvers are typically based on iterative subspace methods, including advanced preconditioners. At the lowest level, large sparse matrix- vector multiplications (spMVM) and vector-vector operations are frequently the most time-consuming building blocks. Most of the available sparse linear (solver) packages were designed in the early 1990s for moderately parallel, homogeneous, and reliable computers (e.g., PETSc [1] or (P)ARPACK [2]) or with a strong focus on object orienta- tion and abstraction (e.g., Anasazi [3]). Numerically intensive kernels are still encapsu- lated in independently developed libraries (see LAMA [4,5] for a recent project), which L. Lopes et al. (Eds.): Euro-Par 2014 Workshop, Part II, LNCS 8806, pp. 578–589, 2014. c  Springer International Publishing Switzerland 2014