Adaptive K-PackCache: Cost-Centric Data Caching in Cloud

2025-09-14

AI Summary

This paper addresses the high data caching cost in cloud-based content delivery networks (CDNs). We propose a dynamic K-ary packing caching strategy that generalizes traditional pairwise (K=2) packing to adaptive, variable-size multi-item joint caching (K ≥ 2). Methodologically, we model access patterns and content correlations to design an online clique formation, merging, and splitting mechanism, jointly optimizing transmission and memory leasing costs. To ensure theoretical guarantees, we introduce approximate clique merging and batched request processing, achieving a bounded competitive ratio. Experiments on Netflix and Spotify datasets demonstrate that our approach reduces total caching cost by 63% and 55%, respectively, compared to state-of-the-art online algorithms, while attaining 85% and 87% of the optimal offline solution’s performance.

Technology Category

• Machine Learning: Online Learning & Bandits
• Data Mining & Knowledge Management: Data Compression
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Cloud, edge and content delivery systems for the Web
• Web Mining and Content Analysis: Content-based information diffusion

Abstract

Recent advances in data analytics have enabled the accurate prediction of user access patterns, giving rise to the idea of packed caching delivering multiple co accessed data items together as a bundle. This improves caching efficiency, as accessing one item often implies the need for others. Prior work has explored only 2 item pairwise packing. In this paper, we extend the concept to general K packing, allowing variable size bundles for improved flexibility and performance. We formulate the K PackCache problem from a content delivery network CDN operator perspective, aiming to minimize total cost comprising two components: transfer cost modeled as a base cost plus a linearly increasing term with the number of items packed, and memory rental cost for caching, which depends on how long and how much is stored. Overpacking increases cost due to low utility, underpacking leads to missed sharing opportunities. We propose an online algorithm, Adaptive K PackCache AKPC, which dynamically forms, merges, and splits data cliques based on user access patterns and content correlation. Our approach supports batch requests, enables approximate clique merging, and offers a formal competitive guarantee. Through extensive evaluation on the Netflix and Spotify datasets, AKPC reduces total cost by up to 63 and 55 percentage over online baselines, respectively, and achieves performance within 15 and 13 percentage of the optimal. This demonstrates its scalability and effectiveness for real world caching systems.

Problem

• Extends caching to variable K-packs for flexible cost optimization
• Minimizes CDN costs balancing transfer and storage expenses
• Dynamically manages data bundles using access pattern correlation

Innovation

• K-PackCache for variable size bundles
• Dynamic clique formation based access patterns
• Cost optimization with transfer and rental components