Vendredi 9 juillet 2021
- Imprimer
- Partager
- Partager sur Facebook
- Share on X
- Partager sur LinkedIn
Vers des méthodes transparentes et parcimonieuses pour l’optimisation automatique des performances
We present a series of descriptions and discussions of various optimizationn methods, from the perspective of performance tuning. We describe heuristics from mathematical optimization, and parametric and nonparametric statistical modeling methods, describing how these surrogate models can be used to minimize an unknown function. We then discuss how the Design of Experiments enables managing the compromise between experimental budget and model quality, establishing a link with Online Learning methods, focusing on parsimony,
progressive model improvement, uncertainty, and robustness, the properties that are most relevant for a method's applicability to autotuning problems.
The key contribution of this thesis is the development of a transparent and parsimonious autotuning approach based on the Design of Experiments, which we apply to diverse problems such as optimizing the configuration of GPU and CPU kernels and finding mixed-precision bit quantization policies for neural networks. We also present a series of empirical evaluations of other methods on autotuning problems from different High Performance Computing domains, such as search heuristics coordinated by a bandit algorithm to optimize the
configuration of compilers for several GPU and FPGA kernels. Although some experimental scenarios eluded the detection and exploitation of search space structure, regardless of the chosen method, we demonstrate how autotuning methods based on the Design of Experiments can aid in interpretable, efficient, and effective code optimization.
Date et Lieu
Vendredi 9 juillet, à 16h00
https://youtu.be/YvfPalv1cq0
Organisé par
Pedro ROCHA BRUEL
Equipe POLARIS - INRIA / LIG
Composition du Jury
Universidade Federal Fluminense (Brésil), examinatrice
Google (France), rapporteur
Argonne National Laboratory (États-Unis), rapporteur
CNRS (France), directeur de thèse
Universidade de São Paulo (Brésil), directeur de thèse
- Imprimer
- Partager
- Partager sur Facebook
- Share on X
- Partager sur LinkedIn