Leveraging Multi-Instance GPUs through moldable task scheduling

Jorge Villarrubia

doi:10.1016/j.jpdc.2025.105128

Leveraging Multi-Instance GPUs through moldable task scheduling

2025-06-06

0

OA

AI

DOI：10.1016/j.jpdc.2025.105128

Original

Original request for help

Save

Abstract

En 中文

• Multi-Instance GPU (MIG) technology is underutilized and has great potential to improve multi-task co-execution through moldable scheduling. • The hypothesis of task monotony, which is traditionally assumed in multi-core moldeable scheduling, is far from being satisfied in MIG, as it is common for tasks to superscale when receiving more resources. • The combination of different heuristics allows to exploit the dynamic reconfiguration of MIG achieving approximations to the optimal scheduling with less than 10% error and remarkable improvements over the use without reconfiguration.

Keywords:

Multi-Instance GPU (MIG)

Moldable Resource Management

Task Scheduling

Leveraging Multi-Instance GPUs through moldable task scheduling

Abstract

Journal

Organization