Accelerating Object Storage for AI/ML with S3 RDMA

Amazon S3 is the de facto standard for object storage—simple, scalable, and accessible via HTTP. However, traditional S3 access via TCP/IP is CPU-intensive and not designed for the low-latency, high-throughput needs of modern GPU workloads. S3 RDMA aims to bridge that gap. S3 RDMA implements S3 object PUT/GET data transfers over RDMA, essentially bypassing the HTTP stack entirely. Data flows back and forth from storage into user memory space using zero-copy mechanisms, aligned with how GPUs prefer to operate for AI/ML workloads. Dell's experience building an S3 RDMA solution for Dell ObjectScale uncovered many valuable benefits compared to traditional S3 access and other RDMA-based storage technologies. These benefits include improvements in client-side data path efficiency, optimized GPU utilization, and high-speed data transfer with fewer control messages. With an eye toward interoperability and standardization, this talk will reflect on this experience for attendees interested in emerging cloud storage and AI/ML data access.