Definition of an asynchronous on-the-fly data compression model on accelerators for HPC

This thesis is related to high-performance computing for numerical simulation of complex physical phenomena.
The CEA provides hardware and software resources to achieve the required computing power.
We have witnessed the advent of accelerators, leading to new problems. In particular, memory management becomes critical for achieving exascale performance as the memory ratio per number of computing units is reducing.
This problem affects all areas requiring a large volume of data. Thus, many aspects of this thesis will be general and of global interest.

This thesis will aim to propose an asynchronous model for making data available through compression/decompression techniques. It should be efficient enough to be used "on the fly" (during computations without slowing them down), allowing memory constraints to be relaxed.
Targeted codes are iterative and sequence different phases. Ideally, all computations will be performed on accelerators, leaving CPU resources unoccupied. The proposed model should take advantage of these specificities. The final goal will be to integrate the work into a representative code to evaluate gains in an industrial context.

Multi-architecture Adaptive Mesh Refinement for multi-material compressible hydrodynamics simulation

CEA DAM is actively developing scientific software in computational fluid mechanics (CFD) for the numerical simulation of compressible and multi-material flows. Such numerical tools requires the use of parallel programming models designed for efficient use of large supercomputers. From the algorithmic point view, the fluid dynamics equations must be discretized and solved using the adaptive mesh refinement (AMR) strategy which allows to reduce the computational cost of such simulations, in particular the number of cells (therefore the memory footprint) and to concentrate the computational work load on the areas of interest (discontinuities, shocks, multi-fluid interfaces, etc. ).

Over the past fifteen years, with the appearance of graphics processors (GPUs), the hardware architectures used in the field of high-performance computing (HPC) have evolved profoundly. This PhD thesis is about designing a parallel implementation of the AMR techniques for the case of multi-material flows with the aim of using as efficiently as possible a GPU-based supercomputer. After required numerical verification and validation process, the developed code will be used to perform numerical simulation of a blast wave and its interaction with surrounding structures.

Top