JASDA: Introducing Job-Aware Scheduling in Scheduler-Driven Job Atomization

TL;DR

JASDA proposes a job-aware, market-inspired scheduling paradigm for MIG-enabled GPUs that decentralizes decision-making by having jobs generate and evaluate multiple subjob variants in response to scheduler-announced time-capacity windows. The scheduler performs per-window clearing using a convex combination of job-side and system-side utilities, aiming to balance utilization, fairness, and temporal responsiveness. Key contributions include the formalization of the bidirectional interaction cycle, a per-window optimal clearing algorithm, a probabilistic safety mechanism for subjobs, and an age-aware prioritization scheme to mitigate starvation. The framework links auction theory with online optimization to enable scalable, explainable, and verifiable scheduling in heterogeneous GPU environments, with future work planned around empirical validation and runtime deployment in R. The approach promises improved utilization and reduced fragmentation by enabling continuous adaptation and transparent feedback between autonomous jobs and the scheduler.

Abstract

The increasing complexity and temporal variability of workloads on MIG-enabled GPUs challenge the scalability of traditional centralized scheduling. Building upon the SJA concept, this paper introduces JASDA-a novel paradigm that extends SJA from a largely centralized scheduling model toward a fully decentralized negotiation process. In JASDA, jobs actively generate and score feasible subjobs in response to scheduler-announced execution windows, while the scheduler performs policy-driven clearing that balances utilization, fairness, and temporal responsiveness. This bidirectional, iterative interaction embeds feedback, calibration, and probabilistic safety directly into the scheduling loop, enabling adaptive and transparent decision-making. By coupling principles from auction theory and online optimization with the temporal granularity of GPU workloads, JASDA provides a scalable foundation for market-aware and fairness-driven resource management-bridging theoretical scheduling models with practical deployment in modern MIG-enabled environments relevant to Artificial Intelligence and Agriculture 4.0.