AI Storage Blueprint
summary
Claims concrete changes: collapsing distributed metadata into a single flat schema yields an order-of-one lookup instead of cross-region, multi-step name-to-address mapping; removing the proxy and pushing logic into a client-side SDK lets clients stream bytes directly to storage nodes; adding a distributed in-memory read plan cache plus a peer-to-peer GPU-host hot object cache delivers about an 80% cache hit rate and a measured ~90% decrease in ingestion times, cutting GPU stalls (previously ~20%) that cost tens of millions per hour.
AI-written summary · about
chapters (11)
- 0:00Welcome
- 1:26Storage matters for AI
- 3:52Maximizing GPU with storage
- 6:42Legacy block storage limits
- 9:51New AI storage architecture
- 11:14Caching bursty reads
- 12:57Research velocity in AI
- 15:45Tiered caching hydration
- 17:32Results and impact
- 18:13Key takeaways
- 20:17Q&A
related talks
24mManipulating Data in an Artificial Intelligence World
37mHow to Accelerate Model Training and Eliminate the I/O bottleneck for the Cloud
31mBuilding a Data Infrastructure for AI/ML
31mGotta Cache 'em All: Scaling AI Workloads With Model Caching in a Hybrid...
24mObject Storage Driven Machine Language Workloads
12mOptimizing AI Data Infrastructure with F5 and
25mOvercoming I/O Bottlenecks in LLM Training with Open-Source Distributed Caching
17mFast Cloud Storage for AI Jobs via Grouped I/O API with Transparent Read/Write...