Bringing Google’s Colossus to PyTorch: Rapid Stor...
summary
Google brings Colossus’ stateful gRPC protocol into the Python stack by adding Rapid buckets to GCSFS and fsspec, bypassing REST and front-end proxies. Reported results: sub-millisecond random reads, 20x faster data access and 6 TB/s aggregate throughput with 20M QPS claimed; appends, hierarchical namespaces and atomic folder renames are native; integration is upstreamed to the GCSFS repo (~30k LOC); no code changes to PyTorch required, only point workloads to a Rapid bucket to realize the improvements.
AI-written summary · about
related talks
04mLightning Talk: Google Cloud Data Services and Cloud Foundry
30mBlazing fast I/O of data in the cloud with Daft Dataframes
56mLarge-Scale Distributed Processing in Google Cloud - 7 lessons...
01h 03mExtreme Performance Architecture
53mWhat I Learned About Going Fast at eBay & Google
39mAccelerating Data Science with RAPIDS
09mInnovating Together for Real-time Impact with Google Cloud
01h 03mGoogle Napa: Scalable Data Warehousing with Robust Query Performance