14 Members of the SCC research groups CSMM, FiNE, RAI and SSPE attended the SIAM CSE conference in Amsterdam. The event was a huge success for both groups, with many interesting talks and interdisciplinary interactions taking place. This experience has helped establish new connections to researchers across fields, which will benefit KIT in the long run.
The FiNE group was present with 8 group members at the SIAM CSE 23 conference. Overall, the group gave 8 talks in 7 sessions. Prof. Hartwig Anzt was also part of the early career panel (PD2) [1] which saw wide community engagement and tried to answer young scientists' concerns in the CSE community, including topics such as research software engineering vs science, conflict management, work-life balance, scientific vs industrial career, and more.
The talks given by FiNE were centered around various aspects of the Ginkgo portable sparse linear algebra framework [2] and reflected the community’s interest in the software. The team provided two overview talks, one by Terry Cojean on research software engineering best practices [3], and one by Hartwig Anzt giving a historic overview of how the Ginkgo library evolved to answer the needs of the various scientific applications [4]. Other presentations detailed new functionalities developed for specific scientific applications.
The first area of focus was the batched iterative sparse solvers and preconditioners where Pratik Nayak presented the general scheme in a session dedicated to this new area of research [5]. Building on this, Yen-Chen Chen presented the specific case for tridiagonal and banded matrices with an implementation outperforming all existing vendor solutions [6]. An external speaker, Paul Lin from LBNL, USA, mentioned the use of Ginkgo’s batched iterative solvers to accelerate the XGC Plasma Fusion application [7].
The next area of focus was mixed-precision functionality, with Yu-Hsiang (Mike) Tsai’s presentation of Ginkgo’s performance portable Algebraic Multigrid (AMG) featuring multiple precision formats for the different levels [8]. Prof. Enrique S. Quintana-Orti from UPV, Spain, presented the performance and energy benefits of leveraging mixed precision at the example of the Ginkgo library’s mixed-precision functionality [9]. Another functionality that received large community interest is the new GPU-resident sparse direct methods presented by Tobias Ribizel, which were developed for the US Exascale Computing Project’s ExaSGD for power grid simulations [10]. Also targeting the acceleration of applications were two talks from the new BMBF ExaSim and PDExa projects by Gregor Olenik and Marcel Koch, respectively. ExaSim focuses on the acceleration of the OpenFOAM CFD software by using Ginkgo as a portable and efficient backend with promising early results [11]. PDExa aims at leveraging Ginkgo’s mixed-precision and batched functionalities to accelerate implicit or semi-implicit time stepping of hyperbolic-parabolic partial differential equations (PDEs) discretized with discontinuous Galerkin (DG) methods [12].
The Computational Science and Mathematical Methods (CSMM) research group was represented by four members, including Gayatri Caklovic, Pia Stammer, Steffen Schotthöfer, and Jasmin Hörter.
Gaya organized a minisymposium on Parallel in time methods [13,14] and presented her work on PInT for hyperbolic nonlinear equations [15]. Pia presented her research on Proton transport for cancer therapy, with a focus on dynamical low-rank approximations [16]. Steffen presented his research on Model order reduction with moment methods, with a focus on neural network-based minimal entropy closures [17].
Charlotte Debus from the junior research group RAI presented results for Predicting ILU(0) Effectiveness for Sparse Matrix Systems via Explainable Machine Learning and René Caspart from the SSPE-team talked about Sustainable Software Development on HPC Systems.
Overall, the conference was a success and we look forward to attending similar events and representing KIT in the future.