MOLINA Roméo
Team : PEQUAN
Arrival date : 10/01/2021
- Sorbonne Université - LIP6
Boîte courrier 169
Couloir 26-00, Étage 3, Bureau 326
4 place Jussieu
75252 PARIS CEDEX 05
FRANCE
Tel: +33 1 44 27 88 76, Romeo.Molina (at) nulllip6.fr
https://lip6.fr/Romeo.Molina
Supervision : Fabienne JÉZÉQUEL
Co-supervision : CHAMONT David - LAFAGE Vincent (IJCLab)
Configuration and control of computing precision, application to low energy gamma radiation measurements
This PhD is part of a collaboration between IJCLab (Orsay, France) and LIP6 (Sorbonne University, Paris, France), and linked to a European collaboration for the measurement of high energy electromagnetic radiation, AGATA. It is funded thanks to the financial support from CNRS through the MITI interdisciplinary program. In the physics of the two infinities, computation is usually performed in double precision, for both simulation or analysis of experimental measurements. The induced loss of speed, which remains moderate on conventional processors, was until now tolerated, thanks to the (partial) security provided by the numerous extra decimals with respect to the required physical accuracy. As a result, research and publications in this domain very rarely deal with floating- point errors. However, the arrival of new computing hardware, especially GPUs, is changing the situation: the speed ratio between single and double precision has no longer the same order of magnitude. It therefore becomes strategic to carry out computation with the minimum required precision, in order to obtain the results much earlier. The PhD student will explore any form of configuration of the numerical precision: manual choice of the floating-point type at compilation or at execution; use of mixed precision in different parts of the application; dynamic change of the precision during iterative algorithms. He (she) will have to propose software methods and techniques favoring this maximumflexibility in numerical precision. These methods and techniques will be applied to the computations carried out in the AGATA experiment. It will of course be necessary to ensure the accuracy of the physical results in all configurations. Numerical validation tools, in particular CADNA, will be used to detect numerical instabilities, and the PhD student will propose algorithmic solutions to compensate for these instabilities, and obtain the best compromise between accuracy and efficiency, on various architectures (CPU or GPU). The numerical validation tools used must be adapted to take into account the targeted computation architectures, the floating-point formats supported by these architectures and the particularities of the data from the AGATA experiment. With regard to precision auto-tuning, this involves extending it to GPUs and widening the choice of accuracy criteria on the results. The numerical validation tools should allow recent formats such as BF16 to be taken into account
2022-2024 Publications
-
2024
- S. Graillat, F. Jézéquel, Th. Mary, R. Molina, D. Mukunoki : “Reduced-Precision and Reduced-Exponent Formats for Accelerating Adaptive Precision Sparse Matrix-Vector Product”, (2024)
- R. Molina, V. Lafage, D. Chamont, F. Jézéquel : “Investigating mixed-precision for AGATA pulse-shape analysis”, EPJ Web of Conferences, vol. CHEP 2023 Proceedings, Norfolk, VI, United States (2024)
- S. Graillat, F. Jézéquel, Th. Mary, R. Molina : “Adaptive precision sparse matrix-vector product and its application to Krylov solvers”, SIAM Journal on Scientific Computing, vol. 46 (1), pp. c30-c56, (Society for Industrial and Applied Mathematics) (2024)
-
2023
- S. Graillat, F. Jézéquel, Th. Mary, R. Molina, D. Mukunoki : “Performance Evaluation of Adaptive-Precision SpMV with Reduced-Precision Formats”, (2023)
- R. Molina, S. Graillat, F. Jézéquel, Th. Mary : “Adaptive Precision Sparse Iterative Solvers”, SIAM Conference on Computational Science and Engineering (CSE23), Amsterdam, Netherlands (2023)
-
2022
- R. Molina, S. Graillat, F. Jézéquel, Th. Mary : “Adaptive Precision Sparse Matrix-Vector Product and its application to Krylov Solvers”, 13es Rencontres Arithmétique de l'Informatique Mathématique (RAIM 2022), Nantes, France (2022)
- R. Molina, S. Graillat, F. Jézéquel, Th. Mary : “Adaptive Precision Sparse Matrix-Vector Product and its Application to Krylov Solvers”, Sparse Days Meeting 2022, Saint-Girons, France (2022)