| Title |
Performance analysis of admission control and scheduling for nonpreemptive workloads under time-varying resource budgets in GPU clusters |
| DOI |
https://doi.org/10.5573/ieie.2026.63.10.92 |
| Keywords |
GPU cluster scheduling; admission control; nonpreemptive workload; time-varying resource budget; performance analysis |
| Abstract |
Under a time-varying GPU resource budget, satisfying the budget constraint when a job is admitted does not guarantee compliance in subsequent intervals because running nonpreemptive jobs retain their allocations. Using 83,140 nonpreemptive jobs from the public Microsoft Philly trace, this study decomposes total instantaneous exceedance in five-minute intervals into a new-admission component and a carryover component from running jobs. It then compares strict first-come-first-served, non-reservation first-come-first-served with subsequent-job search, smallest-GPU-first, and hierarchical budget control under identical samples and budget conditions. Among 66,805 jobs arriving in the evaluation period, all four budget-aware policies produced zero new-admission exceedance under the synthetic time-varying budget B3, whereas the observed exceedance coincided with carryover from running jobs. Every carryover episode began at a budget-reduction slot; 64.3-80.0% were multi-interval episodes, 67.7-80.0% of subsequent exceedance intervals persisted without an additional budget reduction, and the maximum duration was 102,900 seconds. Under the fixed budget B2, smallest-GPU-first achieved the lowest 95th-percentile waiting time of 25,500 seconds and slowdown of 23.08. Instantaneous allocation, budget, and exceedance are measured in GPU allocation units, and cumulative episode exceedance is measured in GPU allocation unit-slots; neither represents power or energy. These findings indicate that budget compliance at admission and instantaneous compliance in subsequent intervals must be evaluated separately, and that the residual allocation of running jobs should be treated as a distinct state variable for nonpreemptive workloads. |