Scholarlink

<p>With the increasing use of cloud services, the GPU as a Service (GPUaaS) market has expanded significantly, supporting applications such as 3D modeling, gaming, and deep learning model training. Cloud providers now offer virtual machines equipped with various types and quantities of GPUs. A significant portion of GPUaaS usage is dedicated to executing tasks related to the training phase of machine learning networks, which require parallel processing and are best suited for GPU-based servers. For cloud service providers to optimally allocate resources and for users to make informed decisions regarding time and cost, accurate predictions of task execution times are essential. However, predicting the duration of machine learning training tasks is complex, as it depends on factors like the characteristics of the neural network and the allocated resources. This research addresses the challenge by proposing the DG-LR method, a novel approach for predicting the training time of neural networks. The DG-LR method operates in two main stages: first, it uses non-deterministic graphs to extract essential features, and then it applies a linear regression algorithm to create a predictive model based on historical data. Evaluation results demonstrate that the DG-LR method improves prediction accuracy while also reducing both training and implementation times compared to existing methods.</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

Performance prediction in training of deep learning applications on GPU-based cloud services

  • 期刊论文

  • 发表日期：

    2026年2月18日

查看全文
  • Arezoo Jahani,
  • Seyed Mohammad Hasani

摘要

With the increasing use of cloud services, the GPU as a Service (GPUaaS) market has expanded significantly, supporting applications such as 3D modeling, gaming, and deep learning model training. Cloud providers now offer virtual machines equipped with various types and quantities of GPUs. A significant portion of GPUaaS usage is dedicated to executing tasks related to the training phase of machine learning networks, which require parallel processing and are best suited for GPU-based servers. For cloud service providers to optimally allocate resources and for users to make informed decisions regarding time and cost, accurate predictions of task execution times are essential. However, predicting the duration of machine learning training tasks is complex, as it depends on factors like the characteristics of the neural network and the allocated resources. This research addresses the challenge by proposing the DG-LR method, a novel approach for predicting the training time of neural networks. The DG-LR method operates in two main stages: first, it uses non-deterministic graphs to extract essential features, and then it applies a linear regression algorithm to create a predictive model based on historical data. Evaluation results demonstrate that the DG-LR method improves prediction accuracy while also reducing both training and implementation times compared to existing methods.