Generating Realistic Wear Distributions for SSDs (Work-in-Progress Report)
summary
A machine-learning aging framework samples real workloads in a simulator, builds lightweight per-block regression models from observed internal SSD activity, and iteratively predicts and injects future block states back into the simulator; testing uses FTLsim on a 256GB SSD with three real workloads and shows predicted wear distributions that match the fully simulated distribution while giving about 33% acceleration versus full simulation; compares against DVS and CML baselines, is modular for integration with Amber and FEMU, and explores the accuracy–efficiency tradeoff.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:20In a nutshell
- 0:55Current art
- 1:31Our approach
- 1:42Challenges
- 2:28System overview
- 3:26Prior work
- 3:48Evaluations
- 4:35Future work
related talks
14mFailureMiner: A Joint Key Decision Mining Scheme for Practical SSD Failure Prediction...
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
20mMSFRD: Mutation Similarity based SSD Failure Rating and Diagnosis for Complex...
17mMulti-view Feature-based SSD Failure Prediction: What, When, and Why
16mReinforcement Learning-Based SLC Cache Technique for Enhancing SSD Write
16mCognitive SSD: A Deep Learning Engine for In-Storage Data Retrieval
56mReduce, Reuse, Recycle: Is this Mantra under attack with Proliferation of Machine(…)
20mTectonic-Shift: A Composite Storage Fabric for Large-Scale ML Training
