Lightweight Adaptive Slack-Aware Scheduling for Cost, Energy, and SLA Trade-offs in Heterogeneous Clouds

Abstract

Cloud task schedulers must balance completion time, monetary cost, energy use, deadline compliance, and resource utilisation, yet many multi-objective methods introduce optimisation overhead or are evaluated under incomparable settings. This paper presents Lightweight Adaptive Slack-Aware scheduling (LASA), a deterministic list scheduler that normalises per-virtual-machine candidate
attributes and changes its weights according to task slack and current load dispersion. LASA requires no training or population-based search. A reproducible Python simulator compares LASA with first-come, first-served (FCFS), round-robin, and greedy cost scheduling on identical heterogeneous workloads of 100–1,000 independent tasks. Twenty paired runs per workload use recorded seeds. At
1,000 tasks, LASA achieved a mean makespan of 931.51 s, cost of $0.2985, estimated energy of 425.85 Wh, no deadline violations, 97.40% mean utilisation, and a normalised objective score of 0.0924. Relative to round-robin, it reducedmakespan by 62.25% and estimated energy by 45.69%. Relative to FCFS, the corresponding reductions were only 1.98% and 1.31%, while scheduling overhead increased from 5.80 to 557.30 ms. Greedy cost remained 1.65% cheaper but produced substantially longer schedules. Sensitivity
results expose a local-versus-global energy-model mismatch.The evidence therefore supports LASA as a lightweight, reproducible compromise in the simulated configurations, not as universal or production-cloud superiority.