Adaptive Multi-Metric Autoscaling for Serverless Platforms
TL;DR
A hybrid congestion-aware autoscaling framework for serverless platforms inspired by TCP slow-start and congestion control principles is proposed, which significantly improves latency performance, scalability, and resource efficiency in serverless environments.
Abstract
Serverless computing has emerged as an efficient paradigm for scalable and cost-effective cloud application deployment through dynamic resource allocation. However, widely used autoscaling mechanisms, particularly CPU-threshold-based approaches such as the Knative Pod Autoscaler (KPA), often exhibit limited adaptability under highly dynamic and bursty workloads, leading to increased latency, SLA violations, scaling instability, and inefficient resource utilization. To address these limitations, this paper proposes a hybrid congestion-aware autoscaling framework for serverless platforms inspired by TCP slow-start and congestion control principles. The proposed method integrates multiple runtime metrics, including workload intensity, CPU utilization, memory usage, request latency, queue congestion, and TCP-inspired congestion window dynamics, into a unified weighted decision model to achieve both responsiveness and stability. The framework is implemented in a Knative-based Kubernetes environment and evaluated using 100,000 requests under 100 concurrent users, with comparisons against Knative KPA and an LSTM-based predictive autoscaling baseline. Experimental results demonstrate that the proposed approach consistently outperforms both baselines across key performance metrics. Specifically, it reduces P50, P90, and P99 response times to 0.29 s, 0.56 s, and 1.12 s, respectively, compared to 0.41 s, 0.88 s, and 2.31 s in KPA. Average response time is reduced to 0.32 s, while throughput increases to 246.80 req/s. In addition, the proposed method improves resource efficiency to 0.87 and decreases total execution time to 405.19 s. Stability is also significantly enhanced, with scaling oscillation rate reduced to 1.5, SLA violation rate reduced to 1.9%, and cold start delay reduced to 0.49 s. Overall, the results confirm that the proposed TCP-inspired hybrid autoscaling approach significantly improves latency performance, scalability, and resource efficiency in serverless environments.