The demand for dedicated computational resources for AI tasks is surging. However, the substantial investment required for high-performance GPUs, coupled with in-creasing market scarcity and extended lead times, poses significant challenges. Simul-taneously, underutilization of these resources is a frequent issue. This paper proposes a solution leveraging cloud-native approaches to utilize unused resources available elsewhere, imposing minimal overhead on the originating cluster. We will present how an “origin” cluster can seamlessly peer with a “donating” cluster, offloading tasks to remote hardware with complete transparency and superior efficiency com-pared to current technologies.

GPU Offloading in MLOps: Navigating the Multicloud Ecosystem for Flexible AI/ML Deployments / Risso, Fulvio; Zangari, Giuseppe. - ELETTRONICO. - (2024), pp. 76-80. (Intervento presentato al convegno Conferenza GARR 2024 - Navigare la complessità. Infrastrutture e competenze digitali per la ricerca tenutosi a Brescia (IT) nel 29-31 maggio 2024) [10.26314/GARR-Conf24-proceedings-13].

GPU Offloading in MLOps: Navigating the Multicloud Ecosystem for Flexible AI/ML Deployments

Fulvio Risso;Giuseppe Zangari
2024

Abstract

The demand for dedicated computational resources for AI tasks is surging. However, the substantial investment required for high-performance GPUs, coupled with in-creasing market scarcity and extended lead times, poses significant challenges. Simul-taneously, underutilization of these resources is a frequent issue. This paper proposes a solution leveraging cloud-native approaches to utilize unused resources available elsewhere, imposing minimal overhead on the originating cluster. We will present how an “origin” cluster can seamlessly peer with a “donating” cluster, offloading tasks to remote hardware with complete transparency and superior efficiency com-pared to current technologies.
2024
978-88-946629-3-1
File in questo prodotto:
File Dimensione Formato  
24GARR - GPU Offloading.pdf

accesso aperto

Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Creative commons
Dimensione 277.76 kB
Formato Adobe PDF
277.76 kB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/3002353