Title

AWS re:Invent 2023 - Accelerate generative AI and ML workloads with AWS storage (STG212)

Summary

Pete Eming from the Amazon S3 team and Jordan Dolman from the Amazon FSx team presented a session on optimizing AWS storage for generative AI and ML workloads.
The session covered the importance of choosing the right AWS storage to pair with AI and ML applications, considering the massive amount of data and the need for high-performance CPU and GPU instances.
They discussed the history of AI and ML, emphasizing the need for machine learning to mimic human decision-making through pattern recognition and logic building.
The presenters highlighted the different perspectives of data scientists and storage admins, aiming to balance both views in the session.
They introduced Amazon FSx for Lustre as a high-performance file system for ML workloads, particularly useful for customers lifting and shifting from on-premises to AWS.
The session also touched on the integration of FSx for Lustre with S3 data lakes, enabling a POSIX-compliant file system interface for S3-stored data.
Pete introduced a new storage class, Amazon S3 Express One Zone, designed for ultra-low latency, high throughput, and high transactions per second, ideal for frequently accessed data.
They announced new features such as the S3 connector for PyTorch, MountPoint for Amazon S3, and local caching for MountPoint, all aimed at improving performance and simplicity for ML workloads.
The session concluded with an invitation to a hands-on lab for further learning.

The presenters emphasized the dichotomy between the perspectives of storage admins and data scientists, suggesting that AWS is working to bridge the gap and provide solutions that cater to both.
The introduction of Amazon FSx for Lustre and its integration with S3 data lakes indicates AWS's commitment to providing high-performance storage solutions that are both familiar to users and offer cloud-native features.
The launch of Amazon S3 Express One Zone reflects AWS's focus on performance, offering a storage class specifically designed for AI and ML workloads that require fast access to large datasets.
The new S3 connector for PyTorch and MountPoint for Amazon S3 demonstrate AWS's efforts to simplify the user experience and improve the performance of ML workloads, regardless of the user's choice of tools or frameworks.
The session highlighted the importance of storage choices in the cost and performance optimization of ML workloads, particularly in the context of data loading and checkpointing.
AWS's strategy appears to be focused on providing a range of storage options and tools that allow customers to optimize their ML workloads based on their specific needs, whether they are lifting and shifting from on-premises environments or building directly on AWS.