PhD Position F/M Efficient Data Compression within the Task-Based Method for High-Performance Applications

The project described in the offer is aimed at bridging the gap between High-Performance Computing (HPC) programming and applications through the innovative use of data compression techniques. The research is grounded in the collaboration between the CAMUS and TONUS teams, bringing together expertise in parallelization, acceleration of computer programs, optimization of compilers, and the development of efficient numerical methods for the approximation of Partial Differential Equations (PDEs), particularly in the context of plasma physics modeling.

The research initiative is inspired by the successful implementation of Wavelet compression in PDE solvers, as demonstrated in Clément Flint's PhD work, and seeks to explore the potential of compression as a generic method in HPC. The project will focus on implementing high-performance compression/decompression kernels, developing new building blocks for integrating compression into task-based applications, investigating the application of compression across various numerical applications, and formulating strategies for the automatic incorporation of compression within the task-based method.

Key aspects of the research include:

• Enhancing compression kernels: Optimizing existing kernels for GPU and CPU, exploring efficient storage formats for non-zero values, and creating mechanisms for partial data manipulation.

• Seamless integration with task-based applications: Adapting runtime systems to manage compressed data objects, potentially modifying the StarPU runtime system to incorporate compression features, and determining the best strategies for handling compression/decompression tasks within the runtime environment.

• Application across numerical domains: The project will test and validate the use of compression in various HPC applications, demonstrating its benefits and identifying any limitations or challenges.

• Automation of compression in HPC: Developing methodologies for the automatic integration of compression techniques into HPC applications, aiming to optimize performance by balancing the overhead of compression with the benefits of reduced data movement and storage requirements.

This research project is highly relevant for advancing HPC applications, particularly in fields requiring the simulation of complex phenomena that are computationally intensive and memory-bound. By developing a framework that effectively integrates compression techniques into HPC programming and applications, the project aims to enable more efficient utilization of modern hybrid and multicore computing architectures, potentially leading to breakthroughs in scientific computing and other areas reliant on large-scale numerical simulations.