SC²S Colloquium - April 10, 2017

From Sccswiki
Revision as of 09:33, 28 March 2017 by Bakhtiar (talk | contribs)
Jump to navigation Jump to search
Date: April 10, 2017
Room: 02.07.023
Time: 3:00 pm, s.t.

Ulrich Huber: A Scalable Mesh Format for Parallel Unstructured Meshes in Scientific Applications

A common goal of all scientific applications is an optimized workflow and good scalability. Therefore the code of such applications as well as the used file formats for initialization need to be optimized, both in time consumption as well as scalability. PXDMF, a mesh format for parallel unstructured meshes providing small file sizes, scalability and small generation and initialization times, is presented in this talk. The format is not tailored to a specific use case, but supports various topologies for cells and can handle additional data, needed by using applications. An initialized mesh is stored in the library within a traversable data structure, by which applications can access all data. Furthermore PXDMF provides all required information, enabling applications to establish a communication structure. Finally PXDMF is compatible to visualization tools, such as ParaView and VisIt and supports homogeneous as well as heterogeneous clusters. All this makes PXDMF applicable to a broad spectrum of scientific applications, using unstructured partitioned meshes, to solve their scenarios.

Nora Hagmeyer: Developing an auto-tuned manycore enabled finite element solver for the numerical modeling of nanoscale light/matter interaction using BOAST

My master thesis/ internship at INRIA, France, was concerned with the auto-tuning of an existing finite element type solver for the numerical modelling of light interaction with nanometer scale structures for the modern accelerator architecture of the Intel Xeon Phi processor, codenamed Knights Landing, by using the BOAST environment. BOAST is an auto-tuning tool, developed at INRIA Grenoble, employing a meta-programming approach for porting kernels to different architectures. From a physical point of view, the program deals with the partial differential system of Maxwell equations in the time domain, coupled to an appropriate differential model for the behaviour of the underlying material. Numerically, these differential equations are solved by a Discontinuous Galerkin approach. This highly local approach, in contrast to classical Finite Element Methods, leads to many very small matrix-vector products, making it highly parallelizable. But, this property also gives rise to the challenge of efficiently using the KNL's 512bit-wide vector registers. In this talk, I will present my approach of using the BOAST tool to create a framework for efficiently applying various SIMD optimization strategies, and I will discuss some results produced by it.