Scholarlink

<p>Serverless computing has gained significant interest as it allows service providers to manage resources efficiently. In serverless systems, providers are highly motivated to use their infrastructure cost-effectively due to the detailed billing modules involved. However, the dynamic nature of serverless workloads presents challenges in maintaining function performance and resource efficiency in parallel processing and distributed computing applications. The existing heuristic scheduling techniques fail to capture the true dynamism and often overlook resource efficiency and application performance. The Single-Agent Deep Q-Learning (SADQL) approach optimizes decisions in simpler environments but struggles to handle the complexity and demand of dynamic serverless environments. A multi-agent Deep Q-Learning (MADQL) framework has been proposed to overcome these limitations. Multiple agents interact with the system that dynamically scales up or down by addressing overutilization and underutilization in serverless applications. Further, experimental analysis has been conducted on a real-world dataset through AWS Lambda. It demonstrates significant improvements, including a 0.96% reduction in average response time, a 1.46% decrease in costs, a 2.43% reduction in energy consumption, and a 15.79% decrease in CPU utilization.</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

MADQL: Multi-Agent Deep Q-Learning for Optimized Job Scheduling in Serverless Computing

  • 期刊论文

  • 发表日期：

    2025年7月29日

查看全文
  • Jasmine Kaur,
  • Inderveer Chana,
  • Anju Bala

摘要

Serverless computing has gained significant interest as it allows service providers to manage resources efficiently. In serverless systems, providers are highly motivated to use their infrastructure cost-effectively due to the detailed billing modules involved. However, the dynamic nature of serverless workloads presents challenges in maintaining function performance and resource efficiency in parallel processing and distributed computing applications. The existing heuristic scheduling techniques fail to capture the true dynamism and often overlook resource efficiency and application performance. The Single-Agent Deep Q-Learning (SADQL) approach optimizes decisions in simpler environments but struggles to handle the complexity and demand of dynamic serverless environments. A multi-agent Deep Q-Learning (MADQL) framework has been proposed to overcome these limitations. Multiple agents interact with the system that dynamically scales up or down by addressing overutilization and underutilization in serverless applications. Further, experimental analysis has been conducted on a real-world dataset through AWS Lambda. It demonstrates significant improvements, including a 0.96% reduction in average response time, a 1.46% decrease in costs, a 2.43% reduction in energy consumption, and a 15.79% decrease in CPU utilization.