Scholarlink

Auto-scaling is essential for real-time resource adjustment in cloud services. Traditional reactive methods often delay in reacting to sudden workload spikes, causing performance degradation. Recent proactive methods alleviate this but still face several obstacles: cold-start issues, improper resource utilization, and lacking a cloud-native AI ecosystem. We propose PheScale, a Transformer-based auto-scaling framework for workload prediction and scaling decision. It employs an Adaptive Sequence Transformation for cold-start, a time-sensitive Gaussian filtering scaler for efficient decisions, and a Libtorch wrapper by Golang to assist invocation of Torch models. Experiments on real-world data demonstrate the effectiveness of PheScale.

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

PheScale: Leveraging Transformer Models for Proactive VM Auto-scaling

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yanqin Zheng,
  • Wang Zhou,
  • Changjian Wang,
  • Jingya Zhang,
  • Wenda Tang,
  • Liang Qi,
  • Tianxiang Ai,
  • Guanghui Li,
  • Bin Yu,
  • Xin Yang

摘要

Auto-scaling is essential for real-time resource adjustment in cloud services. Traditional reactive methods often delay in reacting to sudden workload spikes, causing performance degradation. Recent proactive methods alleviate this but still face several obstacles: cold-start issues, improper resource utilization, and lacking a cloud-native AI ecosystem. We propose PheScale, a Transformer-based auto-scaling framework for workload prediction and scaling decision. It employs an Adaptive Sequence Transformation for cold-start, a time-sensitive Gaussian filtering scaler for efficient decisions, and a Libtorch wrapper by Golang to assist invocation of Torch models. Experiments on real-world data demonstrate the effectiveness of PheScale.