面向云负载预测的自适应知识蒸馏框架
An Adaptive Knowledge Distillation Framework for Cloud Workload Prediction
张缘圆 1丁丁1
作者信息
- 1. 北京交通大学计算机科学与技术学院，北京 100044
- 折叠
摘要
随着云计算的广泛应用，精确的负载预测对大规模数据中心的高效资源管理至关重要。然而，现有预测方法在处理长期预测任务时面临严峻挑战，这主要源于云负载数据固有的动态特性，其序列分布极不稳定，严重阻碍了模型对数据中长期依赖关系和稳定模式的捕获能力。为解决上述问题，提出一种面向云负载预测的自适应知识蒸馏框架（Adaptive Knowledge Distillation for Cloud Workload Prediction，AKD-WP）。首先，设计一个信息融合网络，通过周期性解耦机制分别捕获负载序列中的短期波动与长期趋势，并基于此生成预测结果。其次，引入自适应多教师知识蒸馏策略，集成多个异构教师模型的优势，通过知识迁移增强学生网络对复杂数据特征的建模能力。该策略能够根据各教师模型在当前数据分布下的表现，动态调整其知识迁移的权重，确保学生网络能够高效地从多样化教师模型中汲取有效信息。在真实云数据集上的大量实验表明，所提框架在优于现有最先进方法，充分验证了其在长期云负载预测任务中的有效性和优越性。
Abstract
With the widespread adoption of cloud computing, accurate workload prediction is crucial for efficient resource management in large-scale data centers. However, existing prediction methods face significant challenges when dealing with long-term forecasting tasks, primarily due to the inherent dynamic nature of cloud workload data. The highly unstable distribution of workload sequences severely hinders the model\'s ability to capture long-term dependencies and stable patterns in the data. To address these challenges, this paper proposes an Adaptive Knowledge Distillation Framework for Cloud Workload Prediction (AKD-WP). First, an Information Fusion Network is designed to capture both short-term fluctuations and long-term trends in workload sequences through a periodic decoupling mechanism, based on which the final predictions are generated. Second, an adaptive multi-teacher knowledge distillation strategy is introduced to integrate the strengths of multiple heterogeneous teacher models. Through knowledge transfer, this strategy enhances the student network\'s capability to model complex data features. The approach dynamically adjusts the knowledge transfer weights based on each teacher model\'s current performance on the data distribution, ensuring that the student network effectively assimilates valuable information from diverse teacher models. Extensive experiments on real-world cloud datasets demonstrate that the proposed framework outperforms existing state-of-the-art methods, fully validating its effectiveness and superiority in long-term cloud workload prediction tasks.关键词
云计算/负载预测/知识蒸馏/资源调度Key words
Cloud Computing/Workload Prediction/Knowledge Distillation/Resource Allocation引用本文复制引用
张缘圆,丁丁.面向云负载预测的自适应知识蒸馏框架[EB/OL].(2026-03-24)[2026-03-27].http://www.paper.edu.cn/releasepaper/content/202603-240.学科分类
计算技术、计算机技术
评论