Skip to content
DSML.KZ
Publications

Code · Code Ocean · August 11, 2026

Density-Adaptive Locust Swarm Optimisation with Self-Supervised OLS Warm-Start for Pareto-Optimal Cloud Workflow Scheduling

Mohammed Alaa Fadhil Al-hadeethi
Open via DOI

Abstract

Scientific workflow scheduling on public cloud infrastructure directly determines both time-to-result and operational expenditure for large-scale pipelines in astronomy, genomics, and seismology. The central challenge is that makespan and monetary cost are genuinely competing objectives in heterogeneous virtual machine environments, yet most existing algorithms either optimise a single objective or require a scalar aggregation weight that fixes the cost-time trade-off before search begins. This paper proposes LIWSA and LIWSA-ML, two planning algorithms that maintain an explicit Pareto front without aggregation. LIWSA (Locust-Inspired Workflow Scheduling Algorithm) implements the density-dependent phase polyphenism of desert locusts: each individual's probability of collective versus solitary behaviour is determined by its measured local crowding at every generation, not by elapsed iteration count. LIWSA-ML extends LIWSA with a self-supervised ordinary-least-squares warm-start predictor trained entirely from simulator-generated data, requiring no external libraries, no offline corpus, and no infrastructure beyond a standard JVM, since the predictor runs inside WorkflowSim alongside the search. Experiments across 20 Pegasus benchmark workflows from five scientific families at four scale points (24-1000 tasks), with five independent random seeds, show that LIWSA-ML achieves the highest hypervolume in 17 of the 20 instances (plain LIWSA wins the remaining three by a narrow margin), with an average gain of +9.2% over the state-of-the-art MLEAO and +181.8% over HEFT. On data-intensive workflows, LIWSA-ML simultaneously reduces makespan by up to 78.5% and cost by up to 10.0% relative to HEFT, constituting true Pareto dominance. LIWSA-ML produces 5-29 non-dominated scheduling solutions per run versus 1 for single-objective baselines.

Cloud Computing and Resource ManagementDistributed and Parallel Computing SystemsScheduling and Optimization Algorithmsscientific workflow schedulingcloud computingswarm intelligence

More by Mohammed Alaa Fadhil Al-hadeethi