Learn about Novus storage systems

11/02/2026 Contributors

NetApp Novus is a powerful storage system built for the demands of AI at a very large scale. It extends NetApp's experience in enterprise storage to the extremely high-performance infrastructure needed for training and inference. Understanding its administrative structure and configuration options is crucial for effective management and optimal performance.

Overview

NetApp Novus is a scale-out file storage system solution designed for the throughput and scale needed with modern AI workloads. It joins many independent ONTAP storage clusters into a single namespace that clients see as one file system. Novus is built on parallel NFS (pNFS), which separates metadata from the file data. A dedicated metadata service manages the namespace while the storage clusters serve file data directly to clients in parallel. This removes the throughput ceiling of a single storage cluster which no longer limits how fast the system can deliver data.

A Novus deployment can grow simply by adding storage clusters. Each new cluster adds its full bandwidth and capacity to the namespace, so total throughput increases as the system grows without client remounts or disruptive rebalancing. The metadata service scales separately from the data servers, so you can expand namespace size and metadata performance to fit the workload. Novus is designed to run each storage node at full speed and to treat component failures as routine and non-disruptive at large scale. The primary goal is to keep data flowing to compute even as the system grows to very large sizes.

Clients use a standard NFS client, making Novus simple to adapt to large GPU and compute clusters. And Novus is multi-tenant, so that one deployment can host several isolated customers or workloads, each with its own separate namespaces. It also builds on ONTAP's proven data management and protection. For customers, this means storage that can keep the GPUs busy, lower infrastructure cost by getting more performance from each node, and support very large datasets in a single namespace.

Novus can help organizations that run AI and other data-intensive workloads at scale. This typically includes neoclouds, GPU-as-a-service providers, sovereign clouds, hyperscalers, as well as large enterprises and research organizations running AI training and inference. Novus can benefit any customer using high-performance computing or media processing that needs multiple terabytes per second of throughput and want to scale storage alongside their compute.

System design principles

Novus has been designed and built based on several core principles. These underlying architectural choices drive its performance, scalability, and resilience. This helps customers that need high-performance storage that keeps the GPU busy and reduce infrastructure expenses.

Exascale-ready storage: Novus is designed to be an exascale ready software defined storage foundation stack. It combines horizontally scaled pNFS metadata servers with separate and disaggregated data servers.
Throughput beyond traditional ONTAP cluster limits: Traditionally, a single ONTAP cluster reaches a fixed throughput ceiling as set by its controller count and interconnect bandwidth. AI training and inference need throughput which a single ONTAP cluster cannot reach.
Scale by adding clusters: Novus joins many independent ONTAP clusters, each as small as one HA pair, into a single namespace. Each new cluster contributes its full bandwidth and so the total throughput grows linearly. Clients don't need to remount and the fleet doesn't need rebalancing.
Metadata that scales separately: The metadata service layer and the sharded metadata store can each grow from one to more than 1,000 instances. Metadata capacity and request rate don't depend on how many data nodes are used.
Line-rate data nodes: Each data node should run at full NIC speed for reads and full NVMe speed for writes. Any unused per-node throughput appears as just extra hardware and cost.
Resiliency as the main design constraint: At the scale of 1,000 or more storage nodes, failures are not uncommon. Resiliency is a primary design constraint and not an operational afterthought.

Network File System (NFS) implementation

Novus is a software-defined implementation of the Network File System (NFS) protocol. It relies on NFSv4—specifically the Parallel NFS (pNFS) extension—to split the metadata and data paths. Clients can use standard NFS clients to access the Novus Data Director (NDD) using NFSv4, while accessing the data directly at the ONTAP data servers using NFSv3.

Parallel Network File System (pNFS)

Parallel NFS (pNFS) is a feature introduced with NFSv4.1 that separates file metadata management from file-data access. The NFS server handles metadata operations and provides clients with information about where file data is stored. Clients can then directly access the data at the dedicated storage devices and data servers, rather than sending all reads and writes through the single server.

The NFS server supplies a layout which describes how each file’s data is arranged and how the client can access it. Depending on the layout and storage configuration, a client can communicate with one or more data servers in parallel. This increases data-transfer performance and reduce the data-handling load on the metadata server.

pNFS supports several layout types, including file, block, object, and flexible file. Each type describes a different way of mapping file data to underlying storage and a corresponding way for clients to access it. Novus uses the flexible file layout.

The pNFS flexible file layout is a more recent addition to the layout options. It provides a flexible file-based layout that can describe data stored across multiple data servers and can also support mirrored copies. The client uses the layout information to select an appropriate data-server path for access. This allows deployments to adapt data placement and access behavior to the storage configuration and workload requirements.

Major components

There are two major Novus components which are built on the pNFS standard and NetApp's enterprise storage expertise.

ONTAP data server: The ONTAP data server implements the Novus data plane. It stores file data and serves it to NFS clients. A Novus namespace can spread its file data across several ONTAP clusters, and one ONTAP cluster can serve many namespaces.
Novus Data Director (NDD): The Novus Data Director provides the Novus metadata plane. It manages the namespace metadata, handles client requests for metadata operations, and provides clients with layouts that describe where file data is stored across the ONTAP clusters. Clients interact with the NDD using NFSv4.2 with pNFS, while the actual file data is accessed directly from ONTAP.

Review the Novus architecture for more details about these components and how they work together.

Additional technology and concepts

There are several additional concepts relevant to Novus and the underlying technology.

Namespace: A namespace is a logical structure that organizes and identifies resources, allowing them to be referenced consistently within a defined scope. With NFSv4, the namespace is the hierarchy of files and directories presented to clients; it provides a unified view of exported file systems, which clients can navigate using paths. Each NDD namespace has its own metadata server instance.
Federation: Federation is a technique for connecting independent systems so they can work together while remaining separately managed; it can be applied to identity, application services, networks, databases, and other resources. When used with data providers, federation enables users or applications to access data from multiple sources through a unified interface, while the providers retain control of their data—supporting interoperability and reducing the need to centralize or duplicate information.
Sharding: Sharding is a technique for splitting a database’s data across multiple servers, with each server storing and serving a portion of the data. It improves capacity and performance by distributing storage and the query load.
Border Gateway Protocol: Border Gateway Protocol (BGP) is the routing protocol that exchanges reachability information between autonomous systems—the networks that make up the internet—and selects paths using routing policies and attributes. Novus relies on BGP to support its multi-tenant architecture.
Scalable distributed metadata store: A scalable distributed metadata store is the system managing metadata across multiple nodes, ensuring high availability, fault tolerance, and efficient access. In Novus, the Novus Data Director (NDD) implements a scalable distributed metadata store (or just metadata store) to handle namespace metadata and provide clients with layouts for accessing file data.