|国家预印本平台
首页|基于时序预测的无服务器冷启动优化策略

基于时序预测的无服务器冷启动优化策略

Optimization for Serverless Cold Start Based on Time Series Forecasting

中文摘要英文摘要

无服务器计算是一种新兴的云计算模式,能够降低应用部署和管理复杂度并节约成本,而冷启动问题阻碍了无服务器计算的发展。本文面向无服务器计算中的冷启动优化问题进行研究,基于时间序列分析与预测技术提出了两种冷启动优化策略。自适应预留实例配置策略关注函数调用量时间序列的季节性和趋势性规律,使用时间序列分解和季节性差分自回归移动平均模型来生成函数的预留实例配置。基于深度学习的实例预热与重用策略关注短期的函数调用量变化,使用改进的时间卷积网络深度学习方法和并发期望动态规划算法来预热容器或调整重用策略保活窗口。本文在微软云计算系统公开数据集上进行实验,实验结果表明相比于基准方法,两种优化策略均可以有效降低冷启动率并减少系统资源占用。

Serverless computing is an emerging cloud computing paradigm that reduces the complexity of application deployment and management, while saving costs. However, the problem of cold start hinders the development of serverless computing. This paper focuses on the optimization problem of cold start in serverless computing and proposes two cold start optimization strategies based on time series analysis and forecastingtechniques.The Adaptive Provisioned Concurrency Strategy (APCS) focuses onthe seasonal and trend patterns in the time series of function invocations, using time series decomposition and seasonal autoregressive integrated moving average modelto generate the reserved instance configuration for functions. The Warm-Up and Keep-Alive Strategy(WUKAS)based on deep learningfocuses onshort-term changes of function invocations, using an improved temporal convolutional network and aconcurrency expectation dynamic programming algorithmto warm containersin advance or adjust the window size ofkeep-alivestrategy.Experiments are conducted on a public dataset from Microsoft Azure. The results demonstrate that both optimization strategies effectively reduce the cold start rate and reduce system resource occupancycompared to the baselines.

李国英、宋美娜

计算技术、计算机技术

计算机系统结构无服务器计算时序预测深度学习

computer architectureserverless computingtime series forecastingdeep learning

李国英,宋美娜.基于时序预测的无服务器冷启动优化策略[EB/OL].(2023-12-25)[2025-08-25].http://www.paper.edu.cn/releasepaper/content/202312-70.点此复制

评论