Authors

D. Mattei, I. Smith, A. Ferrari, M. Carbillet

Affiliations

UMR 6525 H. Fizeau (UNS/CNRS/OCA)

Abstract

Post-processing for exoplanet detection using direct imaging requires large data cubes and/or sophisticated signal processing technics. For alt-azimuthal mounts, a projection effect called field rotation makes the potential planet rotate in a known manner on the set of images. For ground based telescopes that use extreme adaptive optics and advanced coronagraphy, technics based on field rotation are already broadly used and still under progress.

In most such technics, for a given initial position of the planet the planet intensity estimate is a linear function of the set of images. However, due to field rotation the modified instrumental response applied is not shift invariant like usual linear filters. Testing all possible initial positions is therefore very time-consuming. To reduce the time process, we propose to deal with each subset of initial positions computed on a different machine using parallelization programming.

In particular, the MOODS algorithm dedicated to the VLT-SPHERE instrument, that estimates jointly the light contributions of the star and the potential exoplanet, is parallelized on the Observatoire de la Côte d’Azur cluster. Different parallelization methods (OpenMP, MPI, Jobs Array) have been elaborated for the initial MOODS code and compared to each other. The one finally chosen splits the initial positions on the processors available by accounting at best for the different constraints of the cluster structure: memory, job submission queues, number of available CPUs, cluster average load.

At the end, a standard set of images is satisfactorily processed in a few hours instead of a few days.