电子学报2025，Vol.53Issue(12)：4408-4428,21.DOI:10.12263/DZXB.20250831
面向云平台弹性伸缩的多头注意力-残差修正容器伸缩行为学习模型
A Hybrid Multi-Head Attention and Residual Correction Model for Elastic Scaling Behavior Learning in Cloud Platforms
摘要
Abstract
The rapid development of cloud platforms and microservice architectures has made elastic scaling a critical mechanism for ensuring both performance and cost efficiency.Although prior studies have advanced workload forecasting and hybrid modeling,most approaches still focus on predicting resource utilization(e.g.,CPU or memory)and then map⁃ping forecasts to scaling actions through threshold rules or controller logic.This forecast-control decoupling amplifies pre⁃diction errors and fails to capture practical mechanisms such as hysteresis,cooldown,and discrete scaling steps,thereby lim⁃iting deployment feasibility.To overcome these limitations,we directly learn scaling behaviors,modeling replica count dy⁃namics as autoscaler control actions.We propose a hybrid model,ARIMA-BiLSTM-MHA,that integrates ARIMA for long-term trend extraction,BiLSTM for residual sequence modeling,multi-head attention for capturing critical temporal depen⁃dencies,and residual correction for improving robustness against bursty and non-stationary workloads.We conduct exten⁃sive experiments on the real-world Alibaba cluster-trace-microservices-v2022 dataset,where we systematically compare our method with baselines including PETformer,SparseTSF,TFEGRU,GRU,Transformer,Seq2Seq-LSTM,Seq2Seq-GRU,Seq2Seq-Transfomer,GRU-LSTM,CNN-LSTM and CNN-LSTM-GRU.Our results demonstrate that our approach consis⁃tently outperforms existing methods,achieving relative improvements of 1.57%~71.56%(MSE),0.72%~46.67%(RMSE),1.57%~59.10%(MAE),1.97%~60.48%(MAPE),and 0.27%~15.70%(R²),with R² reaching up to 0.954 3.Furthermore,we conduct container replica autoscaling experiments based on the DeathStarBench socialNetwork benchmark.We show that the behavior learning-driven strategy,compared with the CPU-threshold HPA strategy,successfully reduces the average rep⁃lica count by approximately 17%while lowering the average P99 latency by 2.11%and effectively suppressing tail-latency spikes during load transitions,thereby significantly mitigating resource over-provisioning.We show that our model can more accurately and stably learn and forecast scaling actions,providing forward-looking decision support for autoscaling in practical cloud environments.关键词
云平台/弹性伸缩/容器伸缩行为/多头注意力机制/残差修正/混合模型/深度学习/时序预测Key words
cloud platforms/elastic scaling/scaling behavior learning/multi-head attention/residual correction/hy⁃brid models/deep learning/time series forecasting分类
信息技术与安全科学引用本文复制引用
赵楠楠,黄志杰,韩淑捷,杨帆,王浩,张佳萌,吴若非,赵彤轩,师雨露,张晓,赵晓南..面向云平台弹性伸缩的多头注意力-残差修正容器伸缩行为学习模型[J].电子学报,2025,53(12):4408-4428,21.基金项目
国家自然科学基金(No.62202382,No.62272394) （No.62202382,No.62272394）
广东省基础与应用基础研究基金(No.2021A1515110080) National Natural Science Foundation of China(No.62202382,No.62272394) （No.2021A1515110080）
Guangdong Basic and Applied Basic Research Foundation(No.2021A1515110080) （No.2021A1515110080）
