Task-Based Performance Portability in HPC - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Books Year : 2021

Task-Based Performance Portability in HPC

Abstract

As HPC hardware continues to evolve and diversify and workloads become more dynamic and complex, applications need to be expressed in a way that facilitates high performance across a range of hardware and situations. The main application code should be platform-independent, malleable and asynchronous with an open, clean, stable and dependable interface between the higher levels of the application, library or programming model and the kernels and software layers tuned for the machine. The platform-independent part should avoid direct references to specific resources and their availability, and instead provide the information needed to optimise behaviour. This paper summarises how task abstraction, which first appeared in the 1990s and is already mainstream in HPC, should be the basis for a composable and dynamic performance-portable interface. It outlines the innovations that are required in the programming model and runtime layers, and highlights the need for a greater degree of trust among application developers in the ability of the underlying software layers to extract full performance. These steps will help realise the vision for performance portability across current and future architectures and problems.
Fichier principal
Vignette du fichier
etp4hpc_wp_tasks_oaumage_2021.pdf (4.59 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-03368013 , version 1 (07-10-2021)

Identifiers

Cite

Olivier Aumage, Paul Carpenter, Siegfried Benkner (Dir.). Task-Based Performance Portability in HPC: Maximising long-term investments in a fast evolving, complex and heterogeneous HPC landscape. European Technology Platform for High Performance Computing (ETP4HPC). 2021, ⟨10.5281/zenodo.5549731⟩. ⟨hal-03368013⟩
83 View
106 Download

Altmetric

Share

Gmail Mastodon Facebook X LinkedIn More