AWS has announced Seekable OCI (SOCI) indexing for Amazon SageMaker AI training and processing jobs. The technology allows SageMaker AI to fetch the files needed to start a container while streaming the remaining image contents in the background.

According to AWS, jobs automatically check whether a specified image has a SOCI index. Indexed images use lazy loading, while images without an index continue using standard downloads. If lazy loading cannot be used, the job falls back to a standard download, AWS says. The company also says adopting SOCI does not require changes to a job’s definition, entry point or data channels.

In AWS benchmarks using eight images ranging from 1.8 GB to 35.4 GB, standard downloads took 48 seconds to 7 minutes 17 seconds. SOCI lazy loading took between 3 and 11 seconds. AWS reports that, in tests with a 13.2 GB image across 13 instance configurations, standard downloads took 3 minutes 16 seconds to 4 minutes 20 seconds, compared with 6 to 8 seconds using lazy loading.

AWS said training performance after the first epoch was the same within run-to-run variation. The first epoch took 3 to 16 seconds longer with lazy loading, but the tested jobs finished 1 to 6 minutes sooner overall.

Users can switch to certain pre-indexed AWS Deep Learning Container variants or create a SOCI index for a custom image. AWS notes that images derived from a Deep Learning Container base require a new index for the final image.