Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robustifying Learning-Augmented Caching Efficiently without Compromising 1-Consistency

Peng Chen, Hailiang Zhao, Jiaji Zhang, Xueyan Tang, Yixuan Wang, Shuiguang Deng

TL;DR

The paper tackles the challenge of making learning-augmented caching robust without sacrificing $1$-consistency. It introduces Guard, a phase-based framework that detects prediction errors and guards eviction candidates to bound regret, achieving a state-of-the-art $2H_{k-1}+2$ robustness with only $O(1)$ per-request overhead while preserving the base algorithm’s time. Guard applies to RB-following algorithms (e.g., BlindOracle, LRB, Parrot), and a complementary ExGuard variant trades some predictor usage for improved smoothness, reaching $ ext{O}igl( rac{ ext{OPT}}{ }igr)$-scaled smoothness bounds. Empirical validation on real datasets (BrightKite, Citi, SPEC CPU2006) and various predictors demonstrates that Guard-based methods provide strong theoretical guarantees and practical performance, closely approaching or surpassing prior robustification approaches. The work highlights lightweight, scalable strategies for integrating machine learning into caching systems with robust performance guarantees.

Abstract

The online caching problem aims to minimize cache misses when serving a sequence of requests under a limited cache size. While naive learning-augmented caching algorithms achieve ideal $1$-consistency, they lack robustness guarantees. Existing robustification methods either sacrifice $1$-consistency or introduce excessive computational overhead. In this paper, we introduce Guard, a lightweight robustification framework that enhances the robustness of a broad class of learning-augmented caching algorithms to $2H_{k-1} + 2$, while preserving their $1$-consistency. Guard achieves the current best-known trade-off between consistency and robustness, with only O(1) additional per-request overhead, thereby maintaining the original time complexity of the base algorithm. Extensive experiments across multiple real-world datasets and prediction models validate the effectiveness of Guard in practice.

Robustifying Learning-Augmented Caching Efficiently without Compromising 1-Consistency

TL;DR

The paper tackles the challenge of making learning-augmented caching robust without sacrificing -consistency. It introduces Guard, a phase-based framework that detects prediction errors and guards eviction candidates to bound regret, achieving a state-of-the-art robustness with only per-request overhead while preserving the base algorithm’s time. Guard applies to RB-following algorithms (e.g., BlindOracle, LRB, Parrot), and a complementary ExGuard variant trades some predictor usage for improved smoothness, reaching -scaled smoothness bounds. Empirical validation on real datasets (BrightKite, Citi, SPEC CPU2006) and various predictors demonstrates that Guard-based methods provide strong theoretical guarantees and practical performance, closely approaching or surpassing prior robustification approaches. The work highlights lightweight, scalable strategies for integrating machine learning into caching systems with robust performance guarantees.

Abstract

The online caching problem aims to minimize cache misses when serving a sequence of requests under a limited cache size. While naive learning-augmented caching algorithms achieve ideal -consistency, they lack robustness guarantees. Existing robustification methods either sacrifice -consistency or introduce excessive computational overhead. In this paper, we introduce Guard, a lightweight robustification framework that enhances the robustness of a broad class of learning-augmented caching algorithms to , while preserving their -consistency. Guard achieves the current best-known trade-off between consistency and robustness, with only O(1) additional per-request overhead, thereby maintaining the original time complexity of the base algorithm. Extensive experiments across multiple real-world datasets and prediction models validate the effectiveness of Guard in practice.

Paper Structure

This paper contains 50 sections, 29 theorems, 20 equations, 10 figures, 14 tables, 10 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Our Contributions
  4. Preliminaries
  5. Algorithm Optimality and $1$-Consistency
  6. RB-Compliant Algorithms
  7. RB-Following Algorithms
  8. Guard: The Robustification Framework
  9. Insights and Overview
  10. Preserving $1$-Consistency
  11. Enhancing Robustness
  12. Applications
  13. Achieving Better Smoothness with ExGuard
  14. Experiments
  15. Experimental Setup
  16. ...and 35 more sections

Key Result

Lemma 3.2

For an RB-compliant algorithm A, we have $| \mathbf{1}_i^\textsc{A} | = | \mathbf{1}_i^* |$ and $\mathbf{0}_i^\textsc{A} = \mathbf{0}_i^*$ after serving each request $r_i$.

Figures (10)

  • Figure 1: Performance with synthetic predictions of NRT on BrightKite.
  • Figure 2: Performance with synthetic predictions of binary labels on BrightKite.
  • Figure 3: Eviction Graph
  • Figure 4: LRU-normalized cost ratios on SPEC CPU2006 Benchmark using the PLECO predictor.
  • Figure 5: LRU-normalized cost ratios on SPEC CPU2006 Benchmark using the POPU predictor.
  • ...and 5 more figures

Theorems & Definitions (61)

  • Definition 3.1
  • Lemma 3.2
  • proof
  • Corollary 3.3
  • Corollary 3.4
  • Proposition 3.5
  • proof
  • Definition 3.6
  • Corollary 3.7
  • Lemma 3.8
  • ...and 51 more