Reinforcement Learning-Based SLC Cache Technique for Enhancing SSD Write
Talkat 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage '20) · 2020-07-13..07-14·
summary
Claims a Q-learning agent that monitors workload and internal SSD state and adjusts SLC cache size and hot/cold separation threshold to maximize write efficiency. States include SLC size, space utilization, previous action, hot-write and update frequencies; reward balances host write cost and recurring GC/migration cost. Q-table fits in 16 KB (5,186 entries). Trace-driven simulation across six workloads shows average write throughput up 77.6% and write amplification down 20.3%, with migration/GC costs reduced ~65% versus prior dynamic schemes.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:12Design Considerations
- 2:13Threshold Separation
- 2:45SLC Management Schemes
- 5:31Reinforcement Learning
- 6:26SLC Cache Manager
- 8:38Reorder Function
- 10:32Comparison
- 11:00Baseline Method
- 11:37SLC Based Scheme
- 12:24Results
- 14:27Conclusion
related talks
15m3L-Cache: Low Overhead and Precise Learning-based Eviction Policy for Caches
17mCacheSlide: Unlocking Cross Position-Aware KV Cache Reuse for Accelerating LLM Serving
11mLearning Cache Replacement with CACHEUS
34mVLLM KV Cache Management: From Cache Reuse To Agent Scenario Optimization
16mGL-Cache: Group-level learning for efficient and high-performance caching
05mGenerating Realistic Wear Distributions for SSDs (Work-in-Progress Report)
19mSolidAttention: Low-Latency SSD-based Serving on Memory-Constrained PCs
16mCognitive SSD: A Deep Learning Engine for In-Storage Data Retrieval