On-Demand Container Partitioning for Distributed ML
summary
A complete build-and-distribution stack that adds a new layer type for big files and a 2DFS two-dimensional filesystem enabling independent allotments, a custom builder, and an extended registry with layer-flattening. Evaluation on 14 real ML models split into 18–82 partitions reports up to 56x faster build times (average 56x) and 25x better caching efficiency while remaining OCI-compliant and runtime-compatible; cost is a per-request flattening overhead (~20 ms) and modest CPU/memory use, and the system works end-to-end for distributed edge deployments.
AI-written summary · about
chapters (11)
- 0:00Container-based distributed ML
- 1:42Containers for model distribution
- 3:27Introducing 2DFS
- 5:11Container image structure
- 9:05Building 2DFS images
- 11:32Distributing 2DFS images
- 13:39Build performance gains
- 15:44Caching with model updates
- 17:072DFS performance trade-offs
- 17:49Distributed ML setup and results
- 18:53Open-source and next steps
related talks
20mTectonic-Shift: A Composite Storage Fabric for Large-Scale ML Training
24mBuilding an ML Application Platform from the Ground Up
28mBuilding Developer-Centric ML Inference Platforms
24mObject Storage Driven Machine Language Workloads
01h 03mScaling up machine learning without tears (and what do programming languages ha...
54mRipped Machine Learning with Distributed Computing
14mContainer-Native ML: Scaling Predictive Customer Segmentation
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...