A lightweight heuristic for cost-efficient IaaS auto-scaling of small-scale web applications

Abstract Pay-per-use Infrastructure-as-a-Service (IaaS) makes web-application hosting affordable for small organisations, yet cost-efficient elasticity remains unsolved for deployments of two to eight virtual machine instances: enterprise auto-scalers demand weeks of traffic history and dozens of tuning parameters, while naive fixed-threshold policies react only after service degradation has begun. This paper proposes the Lightweight Adaptive Scheduling Heuristic (LASH), an O (1) -state two-phase algorithm that minimises hourly IaaS cost subject to a 200 ms P99 latency SLA. Phase 1 applies double exponential smoothing to forecast request rate one VM warm-up horizon ahead; phase 2 selects the minimum-cost instance count while a two-clause minimum-lifetime / billing-aware flag suppresses premature scale-in. LASH is evaluated against four competitive baselines (fixed-threshold, moving-average, recursive-least-squares regression, and AWS Target Tracking) in a trace-driven discrete-time simulation calibrated to AWS EC2 and Azure VM pricing, instance warm-up, and queueing behaviour, across six synthetic load profiles (n = 10 seeded runs per cell; 600 simulated experiments) and, for the AWS EC2 configuration only, the real FIFA World Cup 1998 24-hour production trace (n = 10 replays). In simulation, LASH dominates every baseline on cost across all six profiles and on P99 latency across all but the lowest-CoV profiles, where the regression forecaster _ LR is competitive. The mean cost reduction versus the fixed-threshold baseline is 41. 9 % (BCa 95 % CI 40. 7 %, 43. 1 %, quantifying simulator run-to-run variability rather than deployment uncertainty), with a 23. 7 % P99 latency reduction and a 75. 9 % SLA-violation reduction; against a CPU-target reactive policy modelled on AWS Target Tracking the cost reduction is 13. 5 %. All improvements are statistically significant under the matched-block Friedman test (p < 0. 001, Friedman ^2 = 0. 97) and a corroborating linear mixed-effects model on run-level data. As a simulation study, these results characterise expected behaviour under the modelling assumptions stated in the paper and are not a substitute for measurement on production infrastructure.