📝 SummarXiv

Abstract

Modern GPU clusters, particularly those built on NVIDIA's Multi-Instance GPU (MIG) architecture, often suffer from inefficiencies because jobs are treated as rigid, indivisible blocks that occupy a fixed slice until completion. The reliance on static peak memory estimates exacerbates fragmentation, underutilization, and job rejections. We propose Scheduler-Driven Job Atomization (SJA), a new paradigm that establishes a bidirectional interaction between scheduler and jobs. In SJA, the scheduler advertises available execution gaps, and jobs respond by signaling interest if they can potentially generate a subjob that fits the offered time-capacity window. The scheduler may collect multiple signals for the same slot and, based on its allocation policy (e.g., fairness, efficiency, or SLA priorities), selects which job is granted the slot. Only then does the chosen job materialize a safe, self-contained subjob tailored to that opportunity. Unlike migration or preemption, SJA proactively shapes workloads before execution, thereby avoiding costly state transfers and unpredictable interruptions. It aims to increase GPU utilization, reduce wait times, and minimize migration overhead by aligning jobs with opportunities in real time, ensuring that each admitted subjob is correct by construction. This paper is presented as a concept paper: it introduces the paradigm, defines its building blocks, and outlines future research directions, rather than offering a full experimental evaluation.