AWS announces SOCI-indexed images for faster SageMaker AI job starts
AWS introduces Seekable OCI (SOCI) indexing for Amazon SageMaker AI Training and Processing jobs to reduce container image download overhead. With SOCI indexing, SageMaker AI lazy-loads container images, fetching only the files needed to start the container immediately and streaming the rest in the background as code runs. This allows jobs to begin in seconds rather than minutes. Users can adopt the feature by using pre-indexed AWS Deep Learning Containers or by adding an index to their own images, requiring no changes to job definitions. Benchmarks show this reduces GPU idle time during image pulls by up to 90 percent.
This innovation significantly reduces job start time and infrastructure costs for deep learning workloads, allowing for faster iteration in hyperparameter tuning and experiments without requiring code changes.
Source: AWS Machine Learning Blog · aws.amazon.com