Adaptive Serverless Resource Management via Slot-Survival Prediction and Event-Driven Lifecycle Control

Zeyu Wang ∗, Cuiqianhe Du, Renyue Zhang, Kejian Tong, Qi He, Qiyuan Tian · 2026

Read Adaptive Serverless Resource Management via Slot-Survival Prediction and Event-Driven Lifecycle Control by Zeyu Wang ∗, Cuiqianhe Du, Renyue Zhang, Kejian Tong, Qi He, Qiyuan Tian (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-05465.txt.

Introduction

Wang and five colleagues cut serverless cold starts, the delays when idle functions wake, by predicting how long resource slots survive and managing lifecycles around events. Their dual approach tunes idle times and holds requests smartly, using sliding-window tallies and async handling. Tests across clouds cut cold starts by up to 51.2 percent and nearly double cost efficiency over baselines.

Catalog record

CREATORZeyu Wang ∗, Cuiqianhe Du, Renyue Zhang, Kejian Tong, Qi He, Qiyuan Tian
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29237
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.05465

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Adaptive Serverless Resource Management via Slot-Survival Prediction and Event-Driven Lifecycle Control Zeyu Wang ∗, Cuiqianhe Du, Renyue Zhang, Kejian Tong, Qi He, Qiyuan Tian Abstract Serverless computing eliminates infrastructure management overhead but introduces significant challenges regarding cold start latency and resource utilization. Traditional static resource allocation often leads to inefficiencies under variable workloads, resulting in performance degradation or excessive costs. This paper presents an adaptive engineering framework that optimizes serverless performance through event-driven architecture and probabilistic modeling. We propose a dual-strategy mechanism that dynamically adjusts idle durations and employs an intelligent request waiting strategy based on slot survival predictions. By leveraging sliding window aggregation and asynchronous processing, our system proactively manages resource lifecycles. Experimental results show that our approach reduces cold starts by up to 51.2% and improves cost-efficiency by nearly 2x compared to baseline methods in multi-cloud environments. ∗ Corresponding author Keywords: Serverless Computing, Resource Management, Cold Start, Event-Driven Architecture, Adaptive Optimization, Cloud Infrastructure 1 Introduction Serverless computing has rapidly evolved into a dominant cloud execution model by abstracting infrastructure complexity and offering fine-grained pay-per-use billing [ 12 ] . This paradigm enables developers to deploy event-driven applications without provisioning servers, relying instead on the platform to handle scaling and resource management automatically. However, the transient nature of serverless functions necessitates frequent container initialization, creating a fundamental tension between maintaining high performance and minimizing operational costs [ 1 ] .This tension is exacerbated by the rise of sophisticated, multi-stage AI systems, such as those used for advertising retrieval, which combine multiple large language models and complex data processing pipelines [ 10 ] . Despite these advantages, existing serverless platforms struggle with the cold start problem, where the latency of creating new execution environments significantly impacts response times. Current mitigation strategies often employ static policies or simplistic heuristics that fail to adapt to the stochastic nature of modern production workloads [ 7 ] .Such workloads increasingly include computationally intensive machine learning applications, such as generative models for sequential advertisement recommendation, which are highly sensitive to latency and resource availability [ 8 ] . These rigid approaches frequently result in either excessive resource wastage during idle periods or unacceptable latency spikes during traffic bursts, highlighting the need for more intelligent and adaptive resource management solutions. In this paper, we propose a comprehensive engineering framework that addresses these challenges through adaptive resource management and event-driven architecture. Our solution integrates a dual-strategy optimization mechanism that combines dynamic idle duration adjustment with a probabilistic request waiting strategy. By utilizing real-time pattern analysis and sliding window aggregation, our system predicts resource availability to make proactive allocation decisions. This approach effectively balances the competing demands of latency reduction and cost efficiency in complex serverless deployments. 2 Related Work Recent efforts in serverless optimization have primarily focused on global scheduling strategies and service-level agreement enforcement. These global approaches often employ control-theoretic models or machine learning techniques to anticipate workload fluctuations and adjust capacity accordingly.Xue et al. [ 16 ] demonstrate that integrating a GCN–GRU spatiotemporal graph learner with combinatorial optimization can convert forecasted system risk into dynamic decision variables, reinforcing our move from static serverless heuristics toward prediction-guided, fine-grained resource control.Sophisticated deep learning models, such as transformer-based ensembles, have also demonstrated success in other domains for handling complex, hierarchical pattern recognition tasks like financial event detection [ 5 ] .Gao et al. [ 2 ] showed that coupling ScaNN-based anisotropic-hashing retrieval with the Gemma language model improves retrieval efficiency and contextual response quality in RAG pipelines, providing a concrete latency-sensitive workload case that strengthens ASRM’s motivation for fine-grained, event-driven serverless resource control.In parallel, the application of large language models has demonstrated significant success in other domains for extracting complex semantic features to enhance system performance, such as in recommendation systems [ 13 ] .Similarly, complex hierarchical frameworks are being developed in other domains, like automatic code generation, to unify semantic understanding, syntactic reasoning, and execution-aware feedback for creating robust and reliable software [ 3 ] .These adaptive principles are also applied in other software engineering domains to manage complex, imbalanced tasks; for example, recent fine-tuning frameworks address task imbalance and weak semantic alignment in multilingual code generation [ 17 ] . In parallel, sophisticated multi-expert frameworks have been developed for other complex domains, such as program synthesis, leveraging techniques like reflection-based repair to ensure robust and accurate outcomes [ 4 ] .Long et al. [ 9 ] showed that combining Transformer encoders with InfoNCE loss and knowledge distillation improves semantic discrimination under multilingual, noise-contaminated matching settings, offering a relevant methodological precedent for strengthening ASRM’s workload representation and downstream prediction robustness. To address the limitations of coarse-grained scheduling, subsequent studies have investigated dynamic optimization techniques at the function level. Underlying these high-level strategies are fundamental advancements in isolation and networking technologies. Furthermore, benchmarking studies have provided critical insights into the performance characteristics of function snapshots, guiding the development of more efficient state restoration mechanisms [ 15 ] . 3 Methodology Serverless computing has emerged as a paradigm shift in cloud infrastructure, eliminating the traditional burden of server management while introducing unique engineering challenges in resource optimization and cold start latency. This paper presents a comprehensive engineering solution that addresses the fundamental trade off between resource utilization efficiency and request latency in serverless environments through an adaptive resource management framework integrated with event-driven streaming architecture. Our approach leverages the convergence of Container as a Service technologies with Function-as-a-Service platforms to implement a dual-strategy optimization mechanism: dynamic idle duration adjustment based on real time request pattern analysis and an intelligent request waiting strategy that predicts resource availability windows. The system employs a probabilistic model to forecast slot survival times using sliding window aggregation techniques, enabling proactive resource allocation decisions that significantly reduce cold starts while maintaining cost efficiency. Through integration with modern Event-as-a-Service platforms and edge computing capabilities, our framework addresses the critical challenge of maintaining sub-second response times in high-frequency request scenarios while adapting to burst patterns and variable workloads. The engineering implementation incorporates gRPC-based service interfaces with asynchronous event processing, condition variable synchronization for resource sharing, and adaptive threshold tuning based on temporal request distribution analysis. Our solution demonstrates how practical engineering techniques, combined with mathematical optimization models, can effectively balance the competing demands of performance, cost, and resource efficiency in production serverless deployments, particularly addressing the challenges encountered in multi-cloud environments where vendor-specific optimizations and cold start characteristics vary significantly. 4 Engineering Framework for Adaptive Serverless Optimization The implementation of efficient serverless systems requires addressing fundamental engineering challenges that emerge from the abstraction of infrastructure management. While serverless computing promises automatic scaling and pay-per-use billing, the practical deployment reveals complex trade-offs between cold start penalties, resource utilization, and cost optimization. Our engineering framework tackles these challenges through a combination of event-driven architecture patterns and adaptive resource management strategies that evolved from extensive production deployment experience. 4.1 System Architecture and Event-Driven Integration The core engineering challenge in serverless optimization lies in managing the lifecycle of compute resources without direct control over the underlying infrastructure. Traditional approaches often rely on static configuration parameters, leading to suboptimal performance across varying workload patterns. Our framework implements a gRPC-based service architecture that maintains state through an event-driven model, where each request triggers

Read the complete text