Optimal SSD Management with Predictions
summary
Model: B blocks of n slots, P distinct pages and spare factor alpha; greedy gets write amplification 1/alpha and fragmented-coloring optimality is open. An offline death-time grouping plus phase batching gives write amplification 1+O(1/B). Online with predicted death times and a small RAM simulation yields WA ≈ 1 + O(1/M) + (total L1 prediction error)/sequence length. A lifetime-class algorithm called gladiator needs no extra memory, degrades logarithmically with classes, is prediction-error–optimal, and beats prior heuristics in experiments.
AI-written summary · about
chapters (12)
- 0:00SSD management algorithms
- 1:20SSD lifetime importance
- 2:29SSD algorithm challenges
- 7:02SSD structure and garbage collection
- 10:20Write amplification and online algorithms
- 15:53Interval graphs and offline analysis
- 21:19Grouping pages by lifetime
- 25:29Offline algorithm for death grouping
- 32:18Learning-augmented SSD management
- 35:59Grouping by fixed classes
- 41:19Experimental results and research
- 42:02Q&A
related talks
26mStorage Gardening: Using a Virtualization Layer for Efficient Defragmentation in the WAFL
41mAdvanced Napkin Math: Estimating System Performance from First Principles
42mAdventures in Efficiency and Performance
48mDesign of data-intensive applications
51mDesigning for Performance
01h 04mWe Live in Interesting Times
15mPrinciples and Methodologies for Serial Performance Optimization
01h 09mThere's Treasure Everywhere