A novel flower pollination-based crossover for multi-objective LLM task scheduling problem

Author First name, Last name, Institution

• Mohammad Tubishat, Zayed University, Abu Dhabi CampusFollow
• Ahmed Shuhaiber, Zayed University, Abu Dhabi CampusFollow
• Sharif Naser Makhadmeh, The University of Jordan
• Mohammed Azmi Al-Betar, Ajman University

Document Type

Article

Source of Publication

Knowledge Based Systems

Publication Date

10-9-2026

Abstract

The rapid deployment of Large Language Models (LLMs) across cloud and edge infrastructures has intensified the need for efficient and sustainable inference scheduling. Unlike training, LLM inference involves high-frequency, latency-sensitive workloads that significantly contribute to energy consumption and carbon emissions. This paper formulates the LLM task scheduling problem as a multi-objective optimization model that concurrently minimizes energy usage, carbon footprint, and service-level agreement (SLA) violations across geo-distributed resources. To address the complexity of this NP-hard problem, a novel modified metaheuristic, the Flower Pollination-Based Crossover (FPC) algorithm, which extends the standard Flower Pollination Algorithm by incorporating a crossover mechanism to enhance convergence and solution diversity is proposed. Additionally, a comprehensive optimization framework that supports the comparative evaluation of eight metaheuristic algorithms is developed, including both single-objective and multi-objective variants. To facilitate reproducible evaluation and reflect real-world inference scenarios, three new LLM scheduling datasets, including LLM-1K, LLM-3K, and LLM-10K, are generated. Extensive experiments conducted on three LLM task datasets demonstrate the superior performance of FPC in achieving balanced trade-offs across energy, carbon, and SLA metrics. Results also highlight the strengths and limitations of various optimization strategies under dynamic, environmentally sensitive deployment conditions.

DOI Link

10.1016/j.knosys.2026.116647

ISSN

0950-7051

Publisher

Elsevier BV

Volume

351

Disciplines

Computer Sciences

Keywords

Carbon emissions, Energy efficiency, Flower Pollination-Based Crossover, Large Language Models (LLMs), Multi-objective optimization, SLA violations

Scopus ID

105047656989

Recommended Citation

Tubishat, Mohammad; Shuhaiber, Ahmed; Makhadmeh, Sharif Naser; and Al-Betar, Mohammed Azmi, "A novel flower pollination-based crossover for multi-objective LLM task scheduling problem" (2026). All Works. 8095.
https://zuscholars.zu.ac.ae/works/8095

Indexed in Scopus

yes

Open Access

no