Pasupuleti, Suryachandra Prasad and Wright, Steven A. orcid.org/0000-0001-7133-8533 (2026) Bodhi: Intelligent and Explainable LLM-Assisted HPC Scheduling. In: 33rd IEEE International Conference on High Performance Computing. The 33rd IEEE International Conference on High Performance Computing, 16-19 Dec 2026 International Conference on High Performance Computing. IEEE, IND. (In Press)
Abstract
Job scheduling on High-Performance Computing (HPC) systems involves optimising for different metrics such as utilisation, fairness, wait times, job throughput, and makespan, and is considered an NP-hard problem; often these metrics are conflicting in nature. As software workloads continue to evolve and hardware resources become increasingly complex, there is a growing need for intelligent schedulers capable of making context-aware scheduling decisions. Large language models (LLMs) offer a promising approach by reasoning over system state, multiple feasible scheduling solutions, job queue information, resource availability, and historical scheduling data to optimise multiple scheduling objectives simultaneously. In this paper, we introduce Bodhi, an LLM-assisted scheduling framework built for the Flux scheduler to investigate the use of open-source LLMs for intelligent HPC job scheduling. We evaluate 10 open-source LLMs for the task of HPC job scheduling using both output quality and inference efficiency metrics, demonstrating that quantised Mixture-of-Experts (MoE) models provide the best trade-off between scheduling quality, inference throughput, and energy consumption. While some models achieve strong performance in generating appropriate scheduling actions and coherent reasoning, all evaluated models show comparatively weaker performance in analysing and justifying trade-offs among the multiple feasible scheduling solutions. Finally, we compare two variants of our Bodhi framework, LLM-assisted and utilisation-focused without LLM, against first-come, first-served (FCFS) and EASY backfilling. The utilisation-focused variant achieves the highest resource efficiency but at the cost of wait-time fairness. Bodhi's LLM-assisted scheduling recovers much of this fairness loss, particularly across users, while retaining efficiency gains well above FCFS and EASY.
Metadata
| Item Type: | Proceedings Paper |
|---|---|
| Authors/Creators: |
|
| Copyright, Publisher and Additional Information: | This is an author-produced version of the published paper. Uploaded in accordance with the University’s Research Publications and Open Access policy. |
| Dates: |
|
| Institution: | The University of York |
| Academic Units: | The University of York > Faculty of Sciences (York) > Computer Science (York) |
| Date Deposited: | 04 Nov 2026 12:05 |
| Last Modified: | 04 Nov 2026 12:05 |
| Status: | In Press |
| Publisher: | IEEE |
| Series Name: | International Conference on High Performance Computing |
| Open Archives Initiative ID (OAI ID): | oai:eprints.whiterose.ac.uk:247507 |

CORE (COnnecting REpositories)
CORE (COnnecting REpositories)