AI infrastructure engineers, storage developers, and cloud service providers need fast and secure access to high-capacity file and object storage to support AI workloads.

AI workloads increasingly require high-speed data access for training, fine-tuning, inference context, tool calls, searches, and database lookups. Much of this data lies in files and objects stored both on-premises and in the cloud. Compute accelerators—including GPUs, TPUs, and XPUs—need remote direct memory access (RDMA) that uses NIC- or DPU-accelerated data transfers (like NVIDIA ConnectX NIC or NVIDIA BlueField DPU) and doesn’t copy the data through the server’s CPU-controlled memory. The need for RDMA-accelerated, zero-copy data transfers grows with faster GPU architectures.

Developers building direct access to file and object storage have had to navigate different APIs and protocols across providers. Object storage over RDMA has also lacked a common wire protocol, leaving developers to support provider-specific integrations or rely on traditional access methods.

The expansion of xio-sig and the new Scaled Accelerated Data Access (SCADA) Server SDK offer two ways to build more interoperable storage access for AI workloads.

New developments in xio-sig and SCADA

NVIDIA is expanding xio-sig to include NVIDIA cuObject alongside cuFile, in partnership with Google Cloud and Microsoft. It is also announcing the general availability of cuObject client and server libraries. Developers can use cuObject’s APIs and RDMA wire protocol to build accelerated object-storage applications and servers, while xio-sig provides a path toward enabling the cuObject client to interoperate with any server-side implementation that adheres to the wire protocol.

A new SCADA Server SDK enables storage providers to build SCADA servers that respond to GPU-initiated requests from SCADA clients. IBM has shown interoperability with a prototype that integrates SCADA and IBM Storage Scale. Together with the xio-sig expansion, these efforts give AI developers, storage providers, and cloud services more ways to build and use accelerated storage through shared APIs and protocols.

There is also an emerging class of AI data access involving small, fine-grained I/O requests initiated by accelerators that don’t fit traditional storage media or protocols. Through the NVIDIA Storage-Next initiative, NVIDIA is leading a group of over 40 vendors and customers, including NAND vendors, controller vendors, storage providers, hyperscalers, and application developers, to define how GPU-driven storage should work and turn those advancements into interoperable, open industry standards. The software infrastructure supporting high-throughput, fine-grained, GPU-initiated storage access is SCADA. The FMS blog on open-source cuFile APIs, Storage-Next, and SCADA provides background on these efforts.

cuObject support in xio-sig

The general availability of cuObject libraries provides AI application developers, open source framework developers, storage providers, and storage consumers with standardized methods for accelerated AI data access using both file and object protocols, as shown in Figure 1. AI accelerators can access file and object storage using RDMA, without routing data through the server CPU, enabling higher throughput, lower latency, and reduced CPU utilization for data writes and reads.

Building on its involvement as a maintainer for cuFile in xio-sig, Google Cloud is evaluating expanding its participation for cuObject, reflecting its focus on high-performance cloud file and object storage. Microsoft also looks forward to joining the xio-sig Board to improve interoperability in storage I/O.

xio-sig progress

The repository structure is set up for each of cuFile and cuObject. Headers for cuFile and cuObject, the cuObject wire protocol, and library implementation code for libxFile and xFilekernel code will be shared once the production-ready stack passes conformance tests. The governance documents are currently under review by pending Board Members.

NVIDIA Storage-Next and SCADA expansion

Within the Storage-Next initiative, the new SCADA Server SDK enables storage partners to build servers that receive requests from GPU-based SCADA clients, fulfill them using local or remote storage, and deliver the results over RDMA. The effort also includes the Storage Lender Service and a SCADA command-line utility for configuring and deploying SCADA. These components support storage-provider servers that can interoperate with SCADA clients.

IBM Storage has demonstrated a prototype in which a SCADA client sends requests to their initial version of the Storage Scale SCADA server, built on the new SCADA Server SDK. This shows how storage vendors can collaborate with NVIDIA to build an ecosystem for SCADA’s accelerated, GPU-initiated storage access, paving the way for software infrastructure that could support access to large datasets for applications such as semantic search, recommender systems, and fraud detection.

Get started with cuObject and xio-sig

With xio-sig expanding to include NVIDIA cuObject and cuObject libraries now generally available, storage partners, providers, and consumers can start using cuObject and prepare to contribute to the community’s work on interoperable APIs and protocols for both cuFile and cuObject.

• Watch the xio-sig repository for updates
• NVIDIA FMS blog about xio-sig, cuFile, and Storage-Next
• NVIDIA presentation on SCADA at the 2024 OCP Summit
• Check out and try the OSS Storage Lender Service
• GPUDirect Storage (GDS) documentation
  • Documentation for cuFile APIs
  • Documentation for cuObject
• Link to cuObject Server 2.0.0
• NVIDIA CUDA 13.4 preview downloads

See the notice regarding software product information.