Streamlined Efficiency: Unshackling Kubernetes Image Volumes for Rapid AI Model...
summary
Introduces a concrete system that mounts OCI image volumes without repackaging by storing a reference list (CSV/JSON of source path, mount path, ETag, size) in an OCI layer, having registries return backend blob URLs, and using a block-backed overlay driver (overd / turbo OCI) to translate IO to ranged HTTP reads. Packaging a 22 GB dataset with ~700,000 files drops from ~4 hours to under 2 minutes with Elink; small-file throughput improves substantially and single‑large‑file speed stays comparable. ETags serve as the primary integrity check during streaming.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:04Image startup challenges
- 3:34OCI registries for data
- 6:04OCI storage history
- 8:12Costs of large datasets
- 10:20Registries despite issues
- 11:03Remote snapshot solution
- 12:35Reference lists for streaming
- 14:24Reference lists as OCI artifacts
- 16:15Mount point and loading
- 17:42IO requests with streaming
- 18:50OverlayBD and Turbo OCI
- 20:19Elink performance tests
- 23:03Q&A
related talks
45mOCIv2: Container Images Considered Harmful
21mDADI: Block-Level Image Service for Agile and Elastic Application Deployment
25mStop Re-Downloading Your Container Images: Content-Based Layers with Chunkah
10mLightning Talk: OCI Images for More Than Containers, Jon Johnson, Google
23mBringing Container-Native Simplicity to AI/ML
40mNode-Level Optimization in Large Cluster Powered by Containerd Plugin
27mUnlocking Storage Efficiency: Diagnosing & Optimizing Performance in Scaling Infra
20mFrom Cold Start To Warp Speed: Triton Kernel Caching With OCI Container Images