Scholarlink

<p>With the rapid development of cloud computing and AI technologies, accurate GPU usage prediction has become crucial for resource scheduling in data centers. This paper proposes a Multi-dimensional feature driven GPU usage prediction model based on task and system environment classification that classifies tasks based on their GPU requirements before applying specialized prediction algorithms. We introduce two innovative sub-models: a Transformer model based on dynamic gating mechanism and cross-layer attention head fusion (DGCLA-Transformer) for non-expanding GPU tasks and a neural network model with linear integration of multiple decision trees for expanding scenarios. Experimental results demonstrate remarkable performance, achieving 91% classification accuracy, the RMSE of prediction performance without the need to expand GPU resources is not greater than 0.074; the RMSE of prediction performance is not greater than 0.24 when GPU resources need to be expanded. This work provides a new paradigm for GPU resource management by shifting focus from time-series to task-environment characteristics. The detailed code address of the method is <a href="https://github.com/zsdnr/GPU_prediction">https://github.com/zsdnr/GPU_prediction</a>.</p>

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

Multi-dimensional feature driven GPU usage prediction model based on task and system environment classification

  • 期刊论文

  • 发表日期：

    2025年7月12日

查看全文
  • Chunhui Xu

摘要

With the rapid development of cloud computing and AI technologies, accurate GPU usage prediction has become crucial for resource scheduling in data centers. This paper proposes a Multi-dimensional feature driven GPU usage prediction model based on task and system environment classification that classifies tasks based on their GPU requirements before applying specialized prediction algorithms. We introduce two innovative sub-models: a Transformer model based on dynamic gating mechanism and cross-layer attention head fusion (DGCLA-Transformer) for non-expanding GPU tasks and a neural network model with linear integration of multiple decision trees for expanding scenarios. Experimental results demonstrate remarkable performance, achieving 91% classification accuracy, the RMSE of prediction performance without the need to expand GPU resources is not greater than 0.074; the RMSE of prediction performance is not greater than 0.24 when GPU resources need to be expanded. This work provides a new paradigm for GPU resource management by shifting focus from time-series to task-environment characteristics. The detailed code address of the method is https://github.com/zsdnr/GPU_prediction.