Scholarlink

<p>The loosely-coupled Cloud-Native microservices enable the elastic and scalable application deployment but pose challenges in ensuring Quality of Service (QoS) and optimizing resource allocation due to complex topological dependencies and varied resource requirements. To address these challenges, we propose a Transformer-based Prediction and Resource Adaption Method (TPRAM). Specifically, TPRAM first leverages transformer with its attention mechanism to construct a performance prediction model for the microservices nodes on the critical path, and then employs the Deep Deterministic Policy Gradient (DDPG) reinforcement learning technique to adjust the CPU, memory as well as the I/O disk for each node according to the real-time practical requirements. To validate the effectiveness, we deployed TPRAM on the local cluster, and then simulate the real-world applications based on the benchmark DeathStarBench. The experimental results indicate that compared with the baselines, TPRAM saves at least 40.58% and 15.84% of CPU and memory resources while guaranteeing the end-to-end QoS of the whole system.</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

Transformer-based performance prediction and proactive resource allocation for cloud-native microservices

  • 期刊论文

  • 发表日期：

    2025年8月30日

查看全文
  • Yang Chen,
  • Jia Hao,
  • Yu Peng,
  • Hongyan Xia

摘要

The loosely-coupled Cloud-Native microservices enable the elastic and scalable application deployment but pose challenges in ensuring Quality of Service (QoS) and optimizing resource allocation due to complex topological dependencies and varied resource requirements. To address these challenges, we propose a Transformer-based Prediction and Resource Adaption Method (TPRAM). Specifically, TPRAM first leverages transformer with its attention mechanism to construct a performance prediction model for the microservices nodes on the critical path, and then employs the Deep Deterministic Policy Gradient (DDPG) reinforcement learning technique to adjust the CPU, memory as well as the I/O disk for each node according to the real-time practical requirements. To validate the effectiveness, we deployed TPRAM on the local cluster, and then simulate the real-world applications based on the benchmark DeathStarBench. The experimental results indicate that compared with the baselines, TPRAM saves at least 40.58% and 15.84% of CPU and memory resources while guaranteeing the end-to-end QoS of the whole system.