 |
||
|
||
Home > TERATEC FORUM > Workshop 3
Workshop 03 - 9:00 to 12:30 pm
HPC and computer coding:
a constantly evolving ecosystem
Chaired by Cyril BAUDRY, Scientific IS Architect, EDF R&D and Jean-Albert Vilmer, Regional Director EMEA South, MSC Software Hexagon
© TERATEC 2022
This workshop included 6 presentations showing the close link between hardware, software, tools, libraries, calculation codes and study methodologies.
The first three presentations focused on libraries, compilers, tools and optimization strategies.
The OneAPI software suite is a set of tools for compiling, analyzing, optimizing and debugging. OneAPI includes an implementation of SYCL allowing execution on CPU, GPU, FPGA... from different hardware vendors. Note the presentation of the SYCLomatic tool to allow a migration from CUDA to SYCL.
The MAESTRO middleware has been developed within a European project framework, this middleware allows to optimizing workflows for which very large volumes of data need to be transferred between applications without disk writing. This middleware allows applications to make data available (from providers) that can be used by others (towards consumers) as soon as made available. The result is an optimization of data driven computation time that can be used once available.
The next presentation dealt with computational calculation for biology. The speaker, a computer scientist, presented his strategy for the optimization of computational codes used in biology: knowledge of the scientific domain (biology in this case) and collaborative work are essential to identify bottlenecks and fine-tune the computational codes being used. The need for portability and sustainability in development implies the use of such tools as oneAPI or kokkos.
The three following presentations focus on the calculation codes and study strategies.
The development of ProLB is based on data flow programming and the spread between each team, for the core computing and physical modeling. The core computing team is dedicated to HPC, parallelism... while the modelers use a dedicated language to describe the algorithms. The task of the modelers is thus simplified, with other complex aspects of optimization, porting... being carried out by the core team.
The HEXAGON presentation focuses on acoustics modeling on the one hand and CFD on the other. Acoustic optimization is based on automatic meshing according to frequency, optimization of solvers (use of simple precision...) and parallelism (MPI, multithreading...). Depending on the problem considered, a preliminary study allows considerable gains on the overall time for design.
The last presentation of the workshop presents the strategy used to conduct a multiphase CFD study on a very large number of cells (up to 512 billion cells). As the computation time needed to establish a stationary platform was becoming prohibitive, a nesting doll strategy, with slicing and interpolation, was implemented to go from 1 to 8 and then 64 billion cells. The limits for solvers, libraries, file system... were reached at 512 billion cells.
With the participation of :
|
|
|
|
|
|
|
Introduction
Biographies
