Learning Cache Replacement with CACHEUS
summary
Classifies workloads into four primitives—LFU-friendly, LRU-friendly, scan, churn—and builds CACHEUS, an adaptive ensemble that runs experts ARC, LIRS, LFU plus two new lightweight experts SR-LRU and CR-LFU. Evaluates on 329 real workloads across 6 cache sizes with 17,766 simulations, finds paired t tests that CACHEUS with the new experts is the most consistently performing replacement, and shows augmenting LIRS or ARC with LFU-like experts reduces large miss-rate regressions.
AI-written summary · about
chapters (4)
- 0:00Introduction and setup
- 3:13Introducing CACHEUS concept
- 8:09Experimental results and analysis
- 10:48Conclusion and thanks
related talks
15m3L-Cache: Low Overhead and Precise Learning-based Eviction Policy for Caches
16mReinforcement Learning-Based SLC Cache Technique for Enhancing SSD Write
16mGL-Cache: Group-level learning for efficient and high-performance caching
34mVLLM KV Cache Management: From Cache Reuse To Agent Scenario Optimization
25mTo Cache or Not To Cache? a Tiered KVCache Storage System for Agent ...
20mLearning Relaxed Belady for Content Distribution Network Caching
17mCacheSlide: Unlocking Cross Position-Aware KV Cache Reuse for Accelerating LLM Serving
10mLightning Talk: KV-Cache Centric Inference: Building a State-Aware...