Image Snapshotters for Efficient Container Execution in Particle Physics
Talkat KubeCon + CloudNativeCon Europe 2025 · on Thursday, 2025-04-03· at 17:30· Level 1 | Hall Entrance S10 | Room D·
summary
Benchmarks compare containerd snapshotter plugins—stargz, SOCI, and CVMFS—against overlayFS measuring startup time and bytes pulled; lazy snapshotters start containers almost instantly (overlayFS took ~15s) and typically download <10% of image data, trading faster startup for on‑demand fetch delays during execution. Stargz requires image conversion while SOCI adds an index artifact post‑push to avoid rebuilds. CVMFS serves unpacked images with global deduplication (unpacked.cern.ch, billions of files), supports chunked reads, and all snapshotters fall back to full pulls if needed; many tiny files remain a performance caveat.
AI-written summary · about
chapters (17)
- 0:00Image snapshotters in particle physics
- 1:05Particle physics data challenges
- 5:16Containerizing workloads for experiments
- 6:21Lazy pulling benefits
- 7:25Lazy pulling snapshotter implementations
- 9:55Stargz snapshotter format
- 11:20Stargz snapshotter challenges
- 12:04Sochi snapshotter index artifact
- 13:47CVMFS snapshotter unpacked images
- 14:28What is CVMFS
- 16:12CVMFS in experiments
- 17:34Containers with CVMFS distribution
- 18:16Demo: CVMFS snapshotter
- 21:02Benchmarking lazy pulling vs OverlayFS
- 23:29Performance with complex workloads
- 25:56Lazy pulling advantages
- 26:59Q&A
related talks
30mSpeeding Up Analysis Pipelines with Remote Container Images
34mFast Image Pulls Using IPFS And Opportunistic Caching
43mcontainerd
11mFast and Efficient Container Startup at the Edge via Dependency Scheduling
25mOne Size Fits None: Lightweight Workload-Specialized Confidential Containers
34mLetting Your Containers Dream of Electric Sheep
40mNode-Level Optimization in Large Cluster Powered by Containerd Plugin
37mCache Me If You Can: P2P Image Sharing in Kubernetes with Spegel