Scholarlink

In cloud computing, accurate prediction of cloud workload is essential for efficient resource allocation and Virtual Machine (VM) auto-scaling. Most current research relies on Recurrent Neural Network (RNN) models and attention mechanisms for modeling, while the potential of pure convolutional models has been relatively overlooked. Analysis of cloud workload data reveals significant local features: dramatic short-term change and weak long-term dependence. This paper proposes a pure convolution-based cloud workload prediction model, named Narrow-Deep and Wide-Shallow Convolutional Neural Network (NDWS-CNN). The model features “narrow and deep” local modules arranged in series and “wide and shallow” global modules arranged in parallel to extract both local and global features from cloud workload data. Extensive experiments conducted on public cloud workload datasets (including Alibaba, Google, and Azure) and Traffic datasets demonstrate that the proposed model achieves lower prediction errors and superior performance compared to state-of-the-art (SOTA) methods in the Time Series Forecasting (TSF) domain.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating Narrow-Deep Local and Wide-Shallow Global Convolutional Model for Cloud Workload Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Haohan Liu,
  • Runzi Cui,
  • Yongjin Li,
  • Huaiqing He,
  • Kanghua Hui

摘要

In cloud computing, accurate prediction of cloud workload is essential for efficient resource allocation and Virtual Machine (VM) auto-scaling. Most current research relies on Recurrent Neural Network (RNN) models and attention mechanisms for modeling, while the potential of pure convolutional models has been relatively overlooked. Analysis of cloud workload data reveals significant local features: dramatic short-term change and weak long-term dependence. This paper proposes a pure convolution-based cloud workload prediction model, named Narrow-Deep and Wide-Shallow Convolutional Neural Network (NDWS-CNN). The model features “narrow and deep” local modules arranged in series and “wide and shallow” global modules arranged in parallel to extract both local and global features from cloud workload data. Extensive experiments conducted on public cloud workload datasets (including Alibaba, Google, and Azure) and Traffic datasets demonstrate that the proposed model achieves lower prediction errors and superior performance compared to state-of-the-art (SOTA) methods in the Time Series Forecasting (TSF) domain.