Robustifying Learning-Augmented Caching Efficiently without Compromising 1-Consistency¶
Conference: NeurIPS 2025 arXiv: 2507.16242 Code: Available Area: Algorithms / Learning-Augmented Algorithms Keywords: Online caching, learning-augmented algorithms, robustification, 1-consistency, competitive ratio
TL;DR¶
This paper proposes Guard, a lightweight robustification framework that improves the robustness of a broad class of learning-augmented caching algorithms to \(2H_{k-1}+2\) while preserving 1-consistency and incurring only O(1) additional overhead per request.
Background & Motivation¶
State of the Field¶
Background: The online caching problem requires minimizing cache misses under a cache of size \(k\). Learning-augmented caching algorithms leverage ML predictions to guide eviction decisions, facing an inherent consistency–robustness tradeoff:
Limitations of Prior Work¶
Limitations of Prior Work: 1-consistency: achieves optimality (competitive ratio of 1) when predictions are perfect.
Root Cause¶
Key Challenge: Robustness: performance does not degrade arbitrarily when predictions are severely inaccurate.
Challenges with existing approaches:
Naïve methods (e.g., BlindOracle): 1-consistent but without robustness guarantees — performance can be arbitrarily bad under mispredictions.
Marking methods (e.g., PredictiveMarker): \(4H_k\)-robust but not 1-consistent.
Switching methods (e.g., F&R): \(O(\log k)\)-robust and 1-consistent, but computationally expensive at \(O(n^2 \log k)\), requiring recomputation of the optimal solution.
Core problem: Can robustness be efficiently enhanced while preserving 1-consistency?
Method¶
Overall Architecture¶
Guard is a general robustification framework applicable to any RB-following (relaxed-Belady-following) algorithm. Its design is grounded in three key observations:
- Immediate protection harms consistency: Marking methods protect pages after every request, which precludes 1-consistency.
- Error detection must balance precision and efficiency: The optimal solution recomputation in switching methods is prohibitively expensive.
- Structural properties of RB-compliant algorithms: Optimal eviction behavior exhibits exploitable causal chains.
Key Designs¶
Definition of RB-following algorithms: - Under accurate predictions, pages labeled as 1-pages (Belady label 1) are evicted before 0-pages. - This class includes a broad range of algorithms such as BlindOracle, LRB, and Parrot.
Phase mechanism of Guard: - Execution is divided into phases. - At the start of each phase, all cached pages are marked as "old pages" and added to set \(\mathcal{U}\). - A new phase begins when \(\mathcal{U}\) becomes empty.
Prediction error detection: - A misprediction is detected when a cache miss occurs for a page that has already been evicted in the current phase. - In this case, a random unprotected page from \(\mathcal{U}\) is evicted, and the requested page is guarded. - A guarded page will not be evicted again within the current phase.
Lightweight design: - Only \(\mathcal{U}\) and guard flags need to be maintained. - Implemented with hash tables for O(1) per-request overhead. - No recomputation of the optimal solution is required.
Loss & Training¶
Theoretical analysis framework: - 1-consistency proof (Proposition 4.1): Under perfect predictions, no pages are guarded, and Guard&A behaves identically to A. - Robustness proof (Theorem 4.6): By analyzing the number of old and new page evictions per phase, an upper bound of \(2H_{k-1}+2\) is established.
Key Experimental Results¶
Main Results¶
Average cost ratio on SPEC CPU2006 with PLECO/POPU predictors:
| Algorithm | Prediction | Avg. Cost Ratio | 1-Consistent? | Robust? |
|---|---|---|---|---|
| LRU | None | 1.478 | N/A | N/A |
| Marker | None | 1.404 | No | Yes |
| BlindOracle | NRT | 1.335 | Yes | No |
| PredictiveMarker | NRT | 1.346 | No | Yes |
| LMarker | NRT | 1.335 | No | Yes |
| B.O.&M.D | NRT | 1.294 | No | Yes |
| F&R | NRT | 1.360 | Yes | Yes |
| Guard&B.O. | NRT | 1.226 | Yes | Yes |
Results using the LRB predictor:
| Algorithm | Predictor | Avg. Cost Ratio |
|---|---|---|
| LRU | - | 1.478 |
| Marker | - | 1.394 |
| LRB | LRB | 1.281 |
| MARK0 | LRB | 1.268 |
| Guard&LRB | LRB | 1.171 |
Ablation Study¶
| Dataset / Condition | Guard Effect | Analysis |
|---|---|---|
| BrightKite (synthetic NRT noise) | Low cost ratio maintained as noise increases | Robustness validation |
| BrightKite (binary label flipping) | Outperforms baselines even at high flip rates | Misprediction tolerance |
| SPEC CPU (multiple datasets) | Best or near-best across 13 datasets | Generalizability |
| Varying cache sizes | Effective at both \(k=10\) and \(k=100\) | Scalability |
Key Findings¶
- Guard achieves the state-of-the-art consistency–robustness tradeoff: \((1, 2H_{k-1}+2)\).
- Significant efficiency advantage: O(1) overhead vs. \(O(n^2 \log k)\) for F&R.
- Comprehensive empirical superiority: Lowest cost ratio under both synthetic and real predictors.
- Strong generality: Successfully applied to BlindOracle, LRB, and Parrot — three distinct prediction paradigms.
Highlights & Insights¶
- Elegant phase mechanism: Prediction errors are detected at low cost by tracking whether old pages are re-requested.
- Theory–practice alignment: Theoretical guarantees are thoroughly validated by experiments.
- Plug-and-play design: Can directly augment any RB-following algorithm without modifying its core logic.
- O(1) overhead: Robustness is achieved without increasing asymptotic complexity.
Limitations & Future Work¶
- Whether the robustness bound \(2H_{k-1}+2\) can be further tightened (exploration of theoretical lower bounds).
- Guard is currently limited to RB-following algorithms; extension to a broader algorithm class may be valuable.
- Smoothness optimization via ExGuard shows progress but leaves room for further improvement.
- Real-system performance under very large caches and high concurrency remains to be validated.
- Integration with distributed caching settings has not been explored.
Related Work & Insights¶
- BlindOracle (Lykouris & Vassilvtiskii 2018): The first learning-augmented caching algorithm.
- LMarker (Rohatgi 2020): An improvement over marking-based methods.
- F&R (Sadek & Elias 2024): Achieves 1-consistency and robustness via switching, but is computationally expensive.
- LRB (Song et al. 2020): A practical learning-based method for CDN caching.
- Insight: Simple and principled mechanism design often offers greater practical value than complex optimization.
Rating¶
- Novelty: ⭐⭐⭐⭐ (phase-based error detection mechanism is novel)
- Technical Depth: ⭐⭐⭐⭐⭐ (rigorous and complete theoretical proofs)
- Experimental Thoroughness: ⭐⭐⭐⭐ (validated across multiple datasets and predictors)
- Writing Quality: ⭐⭐⭐⭐⭐ (directly integrable into caching systems)