Design issues for numerical libraries on scalable multicore architectures

Future generations of scalable computers will rely on multicore nodes for a significant portion of overall system performance. At present, most applications and libraries cannot exploit multiple cores beyond running addition MPI processes per node. In this paper we discuss important multicore architecture issues, programming models, algorithms requirements and software design related to effective use of scalable multicore computers. In particular, we focus on important issues for library research and development, making recommendations for how to effectively develop libraries for future scalable computer systems.