SMARTTalk: Teaching SMART Logs to Talk to LLMs
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
Converts 30-day SMART windows into 5-day patches, encodes local trends with CNN patch encoders, clusters embeddings into reusable temporal patterns, and maps patterns to concise phrases fed to LLMs with chain-of-thought prompting and an online pattern memory. Built pipeline produces structured JSON diagnoses, delivers ~50× higher F0.5 than raw prompting, about 25% better health classification than prior SMART methods, bucketed MAE near 10 days, 61% within ±5 days, and adapts to unseen behaviors without retraining.
AI-written summary · about
chapters (19)
- 0:00SMARTLog and LLM integration
- 1:03SSD failure prediction at scale
- 2:30Limitations of smart log models
- 4:58Raw telemetry challenges for LLMs
- 5:40Patch-to-pattern-to-phrase pipeline
- 7:26Offline pattern vocabulary building
- 8:50Online telemetry-to-phrases translation
- 10:16Cross-attribute failure patterns
- 11:40Raw data to chronological stories
- 12:21LLM diagnosis with JSON output
- 13:24Handling unseen patterns
- 14:49Evaluation on Alibaba SSD logs
- 15:51Health status prediction results
- 17:16Urgency prediction for risky drives
- 18:00Explanation and action trustworthiness
- 19:02Ablation study on window and batch sizes
- 20:06Design conclusions and insights
- 20:27Main contributions and applications
- 21:28Closing remarks and artifacts
related talks
10mML-based Log Analysis for Faster Debugging
38mThe Art of Compression: Crafting Insightful Summaries with LLMs
24mLLMs: Powering Lightning Fast Product Innovation
26mHow To Win Friends & Influence LLMs (with Prompt Engineering)
30mTowards Clinically Grounded LLMs
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
05mKeynote: Teaching LLMs New Knowledge
15mFrom Words to Data: Using LLMs to Turn Questions into SQL