Scholarlink

<p>Serverless computing has emerged as a dominant paradigm for deploying scalable, event-driven applications, offering abstraction from infrastructure management. However, despite their advantages, serverless platforms continue to face persistent challenges, including high cold start latency and inefficient resource utilization, particularly under bursty or unpredictable workloads. Traditional mitigation techniques, such as static pre-warming or function-specific prediction models, often lead to excessive resource consumption or suboptimal scheduling decisions. This paper proposes a novel tepid-aware scheduling framework that introduces an intermediate provisioning state—tepid containers—positioned between conventional cold and warm states. These containers are lightweight, runtime-ready environments that can be rapidly customized with dependencies based on predicted function invocation probabilities. We model the scheduling task as a bi-objective optimization problem, jointly minimizing response latency and power consumption under dynamic resource constraints. A heuristic algorithm is developed to assign each function request to one of the three provisioning modes—cold, tepid, or warm—based on system availability and prediction confidence. The proposed method is implemented on Fission.io, and its performance is evaluated across ten diverse scenarios, including synthetic workloads and real-world deployment cases. Experimental results show that tepid-aware scheduling reduces cold start latency by up to 57%, lowers power usage by 31%, and increases throughput by 22% compared to traditional threshold-based scheduling and machine learning (ML)-based baselines. The approach is lightweight, scalable, language-agnostic, and well-suited to multi-tenant and edge computing environments.</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

A Tepid-Aware container scheduler for serverless computing environments targeting throughput and energy efficiency

  • 期刊论文

  • 发表日期：

    2026年6月24日

查看全文
  • Mahrad Hanaforoosh,
  • Mehrdad Ashtiani,
  • Mohammad Abdollahi Azgomi

摘要

Serverless computing has emerged as a dominant paradigm for deploying scalable, event-driven applications, offering abstraction from infrastructure management. However, despite their advantages, serverless platforms continue to face persistent challenges, including high cold start latency and inefficient resource utilization, particularly under bursty or unpredictable workloads. Traditional mitigation techniques, such as static pre-warming or function-specific prediction models, often lead to excessive resource consumption or suboptimal scheduling decisions. This paper proposes a novel tepid-aware scheduling framework that introduces an intermediate provisioning state—tepid containers—positioned between conventional cold and warm states. These containers are lightweight, runtime-ready environments that can be rapidly customized with dependencies based on predicted function invocation probabilities. We model the scheduling task as a bi-objective optimization problem, jointly minimizing response latency and power consumption under dynamic resource constraints. A heuristic algorithm is developed to assign each function request to one of the three provisioning modes—cold, tepid, or warm—based on system availability and prediction confidence. The proposed method is implemented on Fission.io, and its performance is evaluated across ten diverse scenarios, including synthetic workloads and real-world deployment cases. Experimental results show that tepid-aware scheduling reduces cold start latency by up to 57%, lowers power usage by 31%, and increases throughput by 22% compared to traditional threshold-based scheduling and machine learning (ML)-based baselines. The approach is lightweight, scalable, language-agnostic, and well-suited to multi-tenant and edge computing environments.