Research

Predictive Auto-Scaling in Cloud Environments: Evaluating Generalisability, Predictive Accuracy, and Computational Efficiency of Deep Learning Models (LSTM vs. Transformers) using Heterogeneous Cluster Traces

M.Sc. dissertation, University of East London / Aug 2026

Summary

LSTM and Transformer forecasters at the same ~30K-parameter budget, on Alibaba and Google traces. Mostly, neither beat a naive baseline, and which way the data transferred mattered more than the architecture.

Abstract

Reactive cloud auto-scalers respond only after a threshold is breached, and provisioning delay can exceed the window in which a workload spike exhausts capacity. Predictive auto-scaling addresses this lag by forecasting near-term demand ahead of provisioning, but published comparisons of the two dominant forecasting architectures, Long Short-Term Memory (LSTM) networks and Transformers, rarely match parameter budgets or test generalisation beyond a single trace. This dissertation evaluated LSTM and Transformer forecasters at a matched budget of approximately 30,000 parameters, trained and tested on heterogeneous Alibaba and Google cluster traces at 5-, 15-, and 30-minute horizons, against a persistence baseline, across three random seeds. In-provider, only one of twelve trained cells beat persistence in normalised RMSE: the Alibaba-source LSTM at 5 minutes, by 1.6%. Cross-provider zero-shot transfer proved directional rather than symmetric: every model trained on Alibaba and evaluated on Google lost to persistence at all three horizons, by four to ten times the baseline RMSE, while at every horizon at least one model trained on Google and evaluated on Alibaba beat persistence. Median inference latency overlapped almost entirely between architectures, but the Transformer took roughly 1.2 to 3 times longer to train than the LSTM. No single architecture dominated across the conditions tested.