DCSE High Performance Computing Symposium

DCSE High Performance Computing Symposium

In June, we will celebrate the first year of operation of the DelftBlue supercomputer, a unique facility that offers new possibilities to scientists and students with a very low threshold. Besides a summerschool for PhD students from all faculties (see www.aanmelder.nl/143287), we are organizing an afternoon with presentations on the latest developments in High Performance Computing by three international experts, poster presentations and a reception. Participants are cordially invited to present their work on a poster. This event is open to anyone with an interest in HPC and/or numerical linear algebra.

Registration is required to ensure there is enough space.

Speakers and topics:

• Prof. Paolo Bientinesi, director of HPC2N, Umeå, Sweden.Title: Current state of programming languages for linear algebra computations
• Prof. Gerhard Wellein, head of the HPC group at Erlangen Regional Computing Center (RRZE) and DCSE guest professor 2023/24 Tentative title: “Performance Modelling of Large-Scale MPI Applications on Multi-Core Clusters”.
• Prof. Laura Grigori, EPFL Lausanne, Switzerland and DCSE guest professor 2023/24  Title tba.

# Abstracts (others will follow soon)

Current state of programming languages for linear algebra computations

Speaker: Paolo Bientinesi
As matrix computations are the bottleneck in countless workflows and applications, many excellent libraries have been developed. These offer a broad range of computational building blocks, optimized for different aspects such as computing platform, problem size, precision, and matrix properties. However, in spite of the quality of such libraries, we observe that end users are increasingly less likely to use them, at least directly. In fact, realistic workflows are significantly more complex than the functionality of the kernels offered in such libraries, and the intelligent mapping of a workflow onto a set of library kernels is in itself a very challenging task. We refer to this task as the “Linear Algebra Mapping Problem” (LAMP). In practice, users often circumvent the problem by adopting high-level languages such as Matlab, Python, and R, or C++ (in combination with libraries such as Armadillo or Eigen), which offer a convenient high-level syntax for matrix computations, thus increasing user productivity. However, as we show in this talk, these languages are still immature with respect to the solution of the LAMP, and in terms of performance they suffer from vastly suboptimal choices. In a nutshell: High-performance libraries are a necessary, but not a sufficient component for high-performance matrix computations.