Runtime support for automatic placement of workloads on heterogeneous processors - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Runtime support for automatic placement of workloads on heterogeneous processors

Résumé

In order to maintain the continuous growth of the performance of computers while keeping their energy consumption under control, the microchip industry develops architectures capable of processing more and more tasks concurrently. Thus, the upcoming generations of processors may count hundreds of independent processing elements that may differ in their functions and features. As an extensive knowledge of their internals cannot be a prerequisite to their programming, and for the sake of portability, these forthcoming computers necessitate the compilation flow to evolve.In this article, we aim to explore solutions to several challenges that awaits the programming tools of such emerging architectures: by relying on a compilation tool to extract and characterize several levels of execution granularity available in a program, we aim at mapping the detected parallelism on the cores of the system taking into account the costs of communication, managing the execution scheduling in an effective way. Most of that should be amenable without much help or tips from the programmer, and more importantly without over-specializing the source code toward a given architecture. One code should fit for all with a desired performance close enough to hand tuned code so that most of programmer would not have to consider to hand-tune a code.An evaluation is done on an old school Cell-BE processor from the PS3 era to check feasibility on a heterogeneous platform. We show we can obtain performance which can surpass state of the art Cell-BE parallel frameworks.
Fichier non déposé

Dates et versions

hal-04495505 , version 1 (08-03-2024)

Identifiants

Citer

Nicolas Benoit, Stéphane Louise. Runtime support for automatic placement of workloads on heterogeneous processors. 2023 IEEE 16th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC), Dec 2023, Singapore, Singapore. pp.210-217, ⟨10.1109/MCSoC60832.2023.00039⟩. ⟨hal-04495505⟩
7 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More