AWS Deep Learning AMI and Deep Learning Containers now support SOCI snapshotter to reduce container cold start times through lazy loading

AWS Deep Learning AMI and AWS Deep Learning Containers are now enabled with support for SOCI snapshotter and index. SOCI (Seekable OCI) uses layer-based indexing to map file locations within container images, allowing containers to start with only necessary files loaded rather than downloading entire images.

In a benchmark test on a g5.2xlarge instance, standard Docker pull of the vLLM Deep Learning Container (9.72GB compressed, 32.7GB disk usage) took 6m59.099s total time. The same container with SOCI snapshotter in lazy loading mode took 21.125s—fetching only the index and necessary layers while remaining layers load in the background on demand.

Standard Docker pulls of 15–20 GB container images can take 4–6 minutes per instance, causing delays in training jobs, inference endpoints, and automatic scaling events. SOCI lazy loading requires container images to have a SOCI index stored in the registry; AWS Deep Learning Containers with the -soci tag suffix come with SOCI indexes pre-created.