Engineer F/H - Adaptation of scientific codes to a task-based runtime system for exploiting high performance calibration and imaging pipelines in interferometric astronomy

1000scholars; Talence; On-site; EUR 45,000 - 65,000; Full time

Job summary

1000scholars in Talence seeks an Engineer to adapt the DDF-pipeline components to a task-based runtime system, enabling distributed and accelerator parallelism for GPU-based supercomputing. You will analyze, delineate tasks and restructure data flows while preserving scientific correctness across configurations.

You will collaborate with the PLER.RION team and INRIA researchers to extend GPU offloading to high-potential tasks, contributing to research software at the frontier of interferometric

Qualifications

• Strong level in C and Python programming.
• Fluency in Linux environments, Git and continuous integration processes.
• Fluency in parallel, distributed and GPU programming on high-performance platforms.
• Excellent spoken and written English.

Responsibilities

• Adapt calibration and imaging pipeline components to the task-based runtime system.
• Delineate tasks with balanced granularity to maximize scheduling opportunities.
• Adapt data structures for asynchronous data transfers and minimize data dependencies.
• Preserve and verify scientific results after transformation of the application.
• Identify high-potential GPU tasks and adapt routines for GPU offloading.

Job description

Context

Interferometric radio-telescopes rely on the analysis of the interference of signals collected from multiple radio antennas. Collecting, processing, and combining such signals resulting from interferences between antenna pair radio signals enable obtaining astronomical images with a high sensitiveness. The counterpart is the intense computational effort required to extract high quality images from these raw data.

Such a processing is performed by scientific softwares called calibration and imaging pipelines. One of the current worldwide interferometric pipeline for radio-telescopes is DDF-pipeline, constituted from the killMS stage and the DDFacet stage, both developed in particular by Cyril Tasse from the Paris Observatory.

The present Engineer position in computer science is opened within the context of project PLER.RION from a Programme Inria Quadrant (PIQ) France 2030 Grant, starting Fall 2026. At the core of the project is the exploration of leveraging task-based parallel and distributed programming techniques in the DDF-pipeline software components with the aim to accelerate such processing on high-performance heterogeneous computing platforms equipped with GPUs.

The goal of task-based runtime systems is to drive the execution of an application by automatically and dynamically managing the assignment of computations on available processors and accelerators on the one side, and managing related data movements on the other side, such as to minimize completion time. Through clever work scheduling, asynchronous handling of memory transfers, communications and input/output, and by a smart management of replicated data, runtime systems achieve performance gains on many scientific applications compared to ree-form parallelism management. The ambition of Project PLER.RION is to adapt and successfully experiment with such principles on interferometric astronomy on the DDF-pipeline software.

Assignment

Adapting scientific codes to a parallel and distributed task-based runtime system is a technically challenging enterprise. It relies on a deep understanding of the initial structure of the application, of its control and data flows, and on domain-specific characteristics and constraints. It then necessitates a careful and meticulous work to delineate tasks and reorganize data structures, to maximize the optimization potential of the runtime system, while rigorously preserving the algorithmic correctness of the transformed application with respect to the initial application, along all its operating configurations.

The person recruited on this position will have the responsibility to implement and validate this transition of the DDF-pipeline software components to adopt a task-based execution model associating distributed parallelism, multicore parallelism, and accelerator parallelism to enable high performance pipeline runs of the software on GPU-based supercomputing architectures.

Moreover, adopting such an execution model will have the mechanical consequence of enlarging the perimeter of GPU-offloadable processings. The person recruited will thus have the additional responsibility to implement such GPU-offloading for the tasks exhibiting the highest potential for performance gains on GPUs, once the transition of the application to the task-based execution model is completed and validated.

The person recruited will work in tight collaboration with the members of Project PLER.RION, whose Principal Investigator Olivier Aumage, Research at the Inria Research Centre at the University of Bordeaux, has a long experience on runtime systems for high performance computing. The mission will also benefit from collaborations with the Bordeaux Astrophysics Laboratory (LAB), the Paris Observatory, and more widely with the community of the ECLAT Joint Laboratory, headed by Damien Gratadour from LAB, which federates synergies between astrophysics and high performance computing on computational interferometric radioastronomy at the France level.

Main activities

• Adaptation of the calibration and imaging pipeline components to the target task-based runtime system of the project;
• Delineation of tasks of a balanced granularity that maximizes scheduling opportunities and minimizes management overhead;
• Adaptation of data structures to maximize asynchronous data transfers and minimize data dependence constraints;
• Preservation and systematic verification of the scientific result outputs from the transformed application with respect to the reference application;
• Identification of tasks with high potential for performance gains on GPU and adaptation of the corresponding routines to achieve these gains.

Additional activities

• Collaborate with a pluridisciplinary community in astrophysics and computer science, at the frontier between research and production;
• Identify and leverage performance gain opportunities;
• Take into account and harmoniously integrate technical and domain specific constraints.

Skills

Technical skills and level required

• Strong level in C and Python language programming;
• Fluency in software development on Linux environments with a Git version manager and continuous integration processes;
• Fluency in parallel programming, distributed programming and GPU programming on high performance platforms;
• Knowledge of code validation and code verification.

Languages

• Excellent spoken and written English level

Relational skills

• Personal implication and sense of the mission;
• Teamwork;
• Ability at integration and dialog in a plural scientific community.