In the world of data storage, the S3 API has long been the undisputed language of the cloud. It is flexible, scalable, and universal. However, as data-intensive workloads—led by AI and machine learning—began to push the limits of traditional networking, a bottleneck emerged: the TCP/IP stack.
To break this bottleneck, Cloudian recently introduced RDMA (Remote Direct Memory Access) for S3 Compatible Storage. While the initial headlines focused on feeding power-hungry GPUs for AI training, the reality is much more significant. RDMA for S3 isn’t just an “AI feature”—it is a fundamental evolution of object storage transport that should be the new standard for any backend application moving data at scale.
The Technical Shift: From TCP to RDMA
For decades, the S3 API has relied on standard TCP/IP. While reliable, TCP is “chatty” and CPU-intensive. Every data packet must be processed by the host’s operating system kernel, copied from the network interface to the kernel buffer, and then copied again to the application memory. This “context switching” consumes significant CPU cycles and introduces latency.
RDMA for S3 (specifically utilizing RoCE v2) changes the game by implementing “Zero-Copy” data transfer.
The design allows the network interface card (NIC) to transfer data directly from the storage system’s memory into the application’s memory without involving the CPU or the OS kernel. By bypassing the networking stack, we achieve:
- Lower Latency: Dramatic reduction in time-to-first-byte.
- Higher Throughput: Maximizing the wire speed of 100G, 200G, and 400G networks.
- CPU Offloading: Freeing up the host CPU to focus on the actual application logic (like analytics or processing) rather than managing packet flow.
A Universal Benefit: Beyond the AI Lab
While AI was the catalyst, RDMA for S3 provides a performance “free lunch” for virtually any backend infrastructure. If your application moves petabytes of data to over S3 compatible storage, the benefits are immediate:
- Big Data Analytics: Engines like Spark, Presto, and Trino thrive on high-throughput data ingestion. RDMA for S3 allows these engines to query massive datasets with a performance profile that mimics local NVMe drives.
- Backup and Recovery: Modern data protection requires shrinking “Recovery Time Objectives” (RTO). Moving large-scale backups over RDMA ensures that restoration happens at the speed of the hardware, not the speed of the software stack.
- Media & Entertainment: For 8K video editing and high-resolution rendering, RDMA for S3 provides the deterministic, high-bandwidth stream required to prevent frame drops—something traditional S3 compatible storage often struggled with.
- Database Tiering: As databases increasingly use S3 compatible storage for “warm” or “cold” storage, RDMA makes the transition between storage tiers nearly invisible to the end-user.
The Vision: Building an Open, High-Performance Ecosystem
At Cloudian, we believe that for a technology to truly transform the industry, it cannot exist in a vacuum. We have pioneered the implementation of RDMA for S3 because we believe the future of the data center is software-defined and hardware-accelerated.
However, the true power of RDMA for S3 lies in interoperability. We are calling on the broader ecosystem—ISVs, backup vendors, database developers, and fellow storage providers—to join us in adopting RDMA for S3. By implementing this protocol, we can collectively move past the limitations of 20-year-old networking paradigms.
When the “S3 protocol” meets the “RDMA transport,” the result is a storage architecture that is finally fast enough to keep up with the next generation of computing.
Let’s Connect
Cloudian has laid the groundwork. We have the production-ready implementation, and the performance gains are verified. We are looking to collaborate with partners and vendors who want to integrate RDMA for S3 into their own stacks to provide their customers with the ultimate performance tier.
The bottleneck is gone. It’s time to build.
