Tectonic-Shift: A Composite Storage Fabric for Large-Scale ML Training
summary
Composite fabric combines Tectonic HDD capacity with a flash tier called Shift that sits in a consistent‑hash ring and uses Meta’s cashlib for node‑local DRAM/flash caching. An admission layer buckets reads (table/partition/feature), uses application‑aware cache policies that infer future reads from dataset specs plus recent history, and enforces a 100 MB/s write limit. Deployed at petabyte scale since early 2022, it absorbs 1.51-3.28x more IO than LRU baselines and achieves a 29% power reduction.
AI-written summary · about
chapters (8)
- 0:00Introduction to tectonic
- 1:36ML storage challenges
- 3:28Designing composite storage
- 5:51Workload characteristics
- 8:34Architecture of shift
- 11:40Intelligent cache policies
- 15:11Performance evaluation
- 17:55Production impact and summary
related talks
19mOn-Demand Container Partitioning for Distributed ML
12mBehemoth: A Flash-centric Training Accelerator for Extreme-scale DNNs
26mEfficiently serving LLMs at scale.
53mKeynote Address: Lessons Learnt in Trying to Build New Storage Technologies
21mBaleen: ML Admission & Prefetching for Flash Caches
26mSpeed up Your ML Workloads With Kubernetes Powered In-memory Data...
11mEngineering Scalable ML Infrastructure for Modern Apps
15mBuilding Self-Healing Cache Infrastructure