Creación de múltiples instancias lógicas de GPU sobre una única GPU física
Loading...
Official URL
Full text at PDC
Publication date
2023
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
El presente trabajo desarrolla una evaluación experimental de las tecnologías que actualmente permiten compartir un procesador gráfico (GPU) entre distintos procesos. Específicamente, el trabajo profundiza en el estudio de la tecnología Nvidia MIG, introducida con la generación Ampere de procesadores gráficos de Nvidia para, a través de la evaluación de cargas de trabajo con distintas características ejecutadas concurrentemente, evaluar los beneficios que dicha tecnología presenta para reducir la contención en el acceso a recursos compartidos. Los resultados obtenidos revelan que MIG es, actualmente, la única tecnología que permite aislar instancias virtuales dentro de una GPU física, permitiendo la compartición de recursos sin interferencias, y a la vez aumentando la estabilidad en los resultados obtenidos, independientemente de la naturaleza de las cargas de trabajo evaluadas.
This project develops an experimental evaluation of the technologies that currently allow the sharing of a graphics processor (GPU) between different processes. Specifically, the project focuses on the study of Nvidia MIG technology, introduced with the Ampere generation of Nvidia graphics processors. The performed experiments aim to evaluate the improvements to the contention in the access to shared resources, obtainable by MIG, through the evaluation of workloads with different characteristics executed concurrently. The obtained results reveal that MIG is currently the only technology that allows to have completely isolated virtual instances over one physical GPU, allowing resource sharing without interference, and increasing the stability of the obtained results, regardless of the characteristics of the workloads being evaluated.
This project develops an experimental evaluation of the technologies that currently allow the sharing of a graphics processor (GPU) between different processes. Specifically, the project focuses on the study of Nvidia MIG technology, introduced with the Ampere generation of Nvidia graphics processors. The performed experiments aim to evaluate the improvements to the contention in the access to shared resources, obtainable by MIG, through the evaluation of workloads with different characteristics executed concurrently. The obtained results reveal that MIG is currently the only technology that allows to have completely isolated virtual instances over one physical GPU, allowing resource sharing without interference, and increasing the stability of the obtained results, regardless of the characteristics of the workloads being evaluated.
Description
Trabajo de Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática, Curso 2022/2023.
Todas las gráficas y el código que se ha escrito o modificado en este trabajo se puede encontrar en el siguiente repositorio de Github: https://github.com/jucm8/tfg-mig