Multi-Expert Autoscaling Boosts Serverless Efficiency, Cuts Costs

Serverless computing offers benefits like automatic resource management and pay-per-use models, but effective autoscaling remains a significant challenge due to fluctuating workloads, cold-start delays, and inter-function dependencies. This study introduces a new autoscaling framework designed to address these complexities. The proposed framework integrates several key components: it first models serverless applications as directed dependency graphs to identify critical functions using weighted degree centrality. Next, it employs lightweight machine learning models (MLP, LSTM, CNN) for short-term workload forecasting. The outputs from these diverse models are then combined using a performance-weighted probabilistic ensemble, inspired by Bayesian model averaging, to achieve a robust consensus. Finally, a sophisticated controller incorporates cold-start awareness and cost comparisons to make intelligent scaling decisions (scale-up, scale-down, or hold). Experimental results, using real workload traces, demonstrate that this ensemble approach achieves 99.88% prediction accuracy, significantly outperforming other forecasting methods. It also consistently reduces infrastructure costs across various cloud pricing models while successfully maintaining performance targets.