Objective:
Propose innovative solutions to make HPC hardware compliant with real-time constraints in terms of data acquisition, computing pipeline and high level programming models.

The purpose of this project is to merge the needs of HPC and real-time constraints in order to develop solutions that fit both domains, and allow their use for low latency, high throughput, real-time applications, by minimising latency and jitter but also proposing adapted and efficient programming models. These disruptive solutions will be tested onto systems from two different domains: a real-time control of adaptive optics (AO) systems, currently under design for the European Extremely Large Telescope (ELT), and a real-time adaptive beamforming (ABF), used by signal-processing applications (civilian and military).

In the academic and industrial worlds, reducing the latency between the input data and the availability of computing results is more and more required for HPC systems. This requirement means the capability to efficiently feed parallel processors used to address complex computing pipeline within the timeframe. To meet these real-time constraints, the mechanism of acquiring data have to be interleaved with the computations in an optimised way, while keeping three main characteristics of these systems:

  • Flexible: to allow their use on multiple applications and reduce costs,
  • Deterministic: for the control of computation and communication times,
  • High performance: to address the specific needs of compute intensive applications.

Exposing the data-acquisition task to the programmer by including it into the parallel programming model could be the solution to overlap data-transfers with computation. Current programming models lack this capability, as they are mainly functional-driven. Our project is to adapt the use of parallel architectures, such as many-cores, to match real-time constraints, as well as to extend current programming models to support event-driven models in which the computation is triggered by external events.

The primary objectives for this project will be distributed into four main axes of research:

  • Evaluation of different parallel technologies (GPUs, FPGAs, many-cores and DSP fabrics), their corresponding parallel programming models (OpenMP, CUDA, OpenCL, SYCL…) and their limits in the context of embedded real-time applications and adaptation,
  • Innovations relating to parallel programming models to address real-time constraints:
    o Extensions on the language (targeting non HPC experts),
    o Improvement of the parallel execution model,
    o Modification of compiler and run-time mechanisms considering requirements,
  • Experimentation on real-time applications:
    o A signal processing application with size, weight and power constraints (Adaptive BeamForming), commonly used in the signal processing domain, by using signals from multi-directional receivers to build directional (spatial) beams, as a directional sensor could produce,
    o A typical pipeline for the Micado AO system (ELT first light instrument) using these enhanced programming models, including pixel processing, wavefront processing from pyramid wavefront sensor images and wavefront reconstruction including spatial and temporal filtering adapted to ELT constraints,

The performance of these two applications, in terms of performance and latency, will be evaluated on several compute technologies (many-core, GPU, FPGA) through benchmarking.

Project timeline

  • 12 months shared between Thales Research & Technology and l’Observatoire de Paris:
    o For parallel programming training and state of the art (beginning of PhD work),
    o To analyse the existing technologies in the different domains (state of the art)
    o To understand the requirements of the real-time systems,
  • 12 months (2 secondments of 6 months) at Barcelona Supercomputing Center for:
    o Compilation principles, methods and tools training,
    o Evaluation of parallel programming extensions (with special interest on OpenMP and CUDA),
    o Development of compiler and runtime techniques to support the proposed extensions.
  • 12 months shared between Thales Research & Technology and l’Observatoire de Paris to:
    o Apply the techniques to the identified systems (supported by TRT team and the engineering team of the Micado project),
    o Finalise the analysis, complete the research and write the thesis.

Fichier PDF
Télécharger le PDF