| 张丁杰* **,武庆华*,谢高岗***.基于贝叶斯优化的微服务资源弹性分配机制[J].高技术通讯(中文),2026,36(3):268~278 |
| 基于贝叶斯优化的微服务资源弹性分配机制 |
| Bayesian optimization based auto-scaling of resource allocation mechanism for microservices |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 03. 005 |
| 中文关键词: 微服务； 资源管理； 弹性伸缩； 贝叶斯优化 |
| 英文关键词: microservice, resource allocation, autoscaling, Bayesian optimization |
| 基金项目: |
| 作者 | 单位 | | 张丁杰* ** | (*中国科学院计算技术研究所北京 100190)
(**中国科学院大学北京 100049)
(***中国科学院计算机网络信息中心北京 100083) | | 武庆华* | | | 谢高岗*** | |
|
| 摘要点击次数: 39 |
| 全文下载次数: 36 |
| 中文摘要: |
| 数据中心网络中的许多延迟敏感型应用程序采用微服务架构。为了同时满足微服务应用延迟需求和提升微服务资源利用率，本文提出了微服务资源弹性分配机制BScaler。BScaler将微服务资源分配问题建模为黑盒函数优化问题，利用贝叶斯优化求解，以获取满足延迟目标的最少资源分配方案。为加速贝叶斯优化求解过程，构建基于图卷积神经网络（graph convolutional network，GCN）的微服务延迟预测模型，替代贝叶斯优化中的采样过程。当应用负载变化时，BScaler根据当前负载、微服务拓扑关系，生成满足延迟需求下每个微服务的资源配置方案，触发弹性伸缩动作，从而保障系统服务水平目标和减少数据中心资源浪费。基于仿真和真实环境的大量实验结果表明，与当前广泛采用的Kubernetes 水平弹性伸缩机制相比，在请求延迟满足服务水平目标的条件下，BScaler使微服务副本数量减少了3.9%~6.2%；与基于机器学习的资源分配机制DScaler相比，BScaler减少了5.6%的微服务副本数量。 |
| 英文摘要: |
| Many latency-sensitive applications in data center networks adopt microservice architecture. To simultaneously meet the latency requirements of microservice applications while improving resource utilization, this paper proposes BScaler, a mechanism for elastic resource allocation in microservices. BScaler models the microservice resource allocation problem as a black-box function optimization problem and employs Bayesian optimization to determine the minimal resource allocation solution that satisfies latency objectives. To accelerate the Bayesian optimization process, a graph convolutional network (GCN)-based microservice latency prediction model is constructed to replace the sampling process in Bayesian optimization. When workload fluctuates, BScaler generates resource configuration plans for all microservices based on current load patterns, request types, and microservice topology, triggering autoscaling actions to maintain service level objectives(SLO) while reducing resource waste. Extensive experimental results conducted in both simulation and real-world environments demonstrate that, compared to the widely-adopted Kubernetes HPA elastic scaling mechanism, BScaler reduces the number of microservice replicas by 3.9%~6.2% while meeting request latency SLOs. Compared to DScaler, a machine learning-based resource allocation mechanism, BScaler reduces the number of microservice replicas by 5.6%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
