Amazon Web Services (AWS) has announced a significant enhancement to Amazon SageMaker AI Training and Processing jobs through the introduction of Seekable OCI (SOCI) indexing. This technology eliminates the need to download entire container images before job startup, enabling SageMaker to lazy-load images by fetching only essential files initially and streaming the remainder in the background.
Traditional container downloads can take several minutes, especially for large deep learning images exceeding 10 GB, resulting in costly idle GPU time and longer iteration cycles. SOCI indexing addresses this bottleneck by providing a separate OCI-compliant index artifact stored alongside the container image in Amazon Elastic Container Registry (ECR), allowing immediate container startup.
Benchmarks demonstrate that SOCI-indexed images reduce download times dramatically—from minutes down to just seconds—yielding speedups of up to 35x for the largest images tested. This improvement accelerates job startups without impacting training performance and requires no changes to job definitions, code entry points, or data pipelines.
AWS offers pre-indexed SOCI variants for popular Deep Learning Containers, including PyTorch and TensorFlow, and provides tools for creating SOCI indexes for custom images. Users can also automate index creation on image pushes using AWS CloudFormation templates.
By adopting SOCI indexing, SageMaker users can reduce costs, shorten machine learning iteration cycles, and enhance productivity while maintaining existing workflows unchanged.
Learn more and start using SOCI-indexed images in your SageMaker AI jobs by visiting the original AWS announcement: