GL-Cache: Group-level learning for efficient and high-performance caching
summary
Groups objects by insertion-time grouping and trains a model on group snapshots (seven features, using gradient-boosted trees) so inference and state are amortized across tens–hundreds of objects. Implemented on an open-source production cache and evaluated on 118 production block I/O and CDN traces, GL-Cache raises hit ratio and throughput: compared to LRB it reports 228× throughput vs LRB and +7% hit ratio on average (P90 +25%), and versus the best prior learned caches it claims 64% higher throughput and +3% average hit ratio (P90 +13%).
AI-written summary · about
chapters (12)
- 0:00Start
- 0:30What location are they going?
- 1:25Introduction
- 6:51GL-Cache architecture
- 7:26Design decision
- 7:48How does GL-Cache group objects Insertion-time-based grouping
- 8:52What does GL-Cache learn A new utility function
- 11:56How does GL-Cache use the model Inference
- 13:00Evaluation setup
- 13:58Efficiency
- 14:33Throughput
- 15:01Summary
related talks
15m3L-Cache: Low Overhead and Precise Learning-based Eviction Policy for Caches
20mLearning Relaxed Belady for Content Distribution Network Caching
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
31mGotta Cache 'em All: Scaling AI Workloads With Model Caching in a Hybrid...
25mOvercoming I/O Bottlenecks in LLM Training with Open-Source Distributed Caching
14mHyCache: Hybrid Caching for Accelerating DNN Input Preprocessing Pipelines
11mLearning Cache Replacement with CACHEUS
16mReinforcement Learning-Based SLC Cache Technique for Enhancing SSD Write