
GPFS vs Lustre: IBM Storage Scale and Lustre Compared
20 November, 2026
Technical Guide
GPFS and Lustre are usually the first two names on the list when an HPC cluster outgrows a single NFS server. This article compares them using the vendors’ own documentation, then shows where BeeGFS fits. Facts are current as of November 2026.
Short answer
GPFS, now IBM Storage Scale, is a commercial, capacity-licensed cluster file system for Linux, AIX and Windows. It includes policy-based tiering, AFM for sharing and replicating data between clusters, and NFS, SMB and S3 access through protocol nodes. Lustre is open source under GPL 2.0 and runs on Linux. It separates metadata servers (MDS/MDT) from data servers (OSS/OST) and is best known from large HPC centres.
Choose Storage Scale for mixed operating systems, protocol access to the same data and a single vendor for software and support. Choose Lustre for large Linux clusters when you want open-source software and have staff who can run it. BeeGFS also uses separate metadata and storage services, and is worth shortlisting for Linux HPC and AI clusters.
If your question is really NFS vs GPFS, see our NFS vs parallel file system comparison. NFS exports a file system from one server, while GPFS and Lustre spread data across many.
What GPFS (IBM Storage Scale) is
IBM’s support pages list the product as IBM Storage Scale, formerly Spectrum Scale and GPFS. The product overview calls it “a clustered file system that provides concurrent access to a single file system” from multiple nodes. Nodes can be SAN-attached, network-attached, a mix of both, or shared-nothing, and one cluster can mix AIX, Linux and Windows Server nodes. Each node runs administration commands, a kernel extension and a multithreaded daemon.
The current product editions are:
- Data Access Edition: the base functions, including ILM and AFM.
- Data Management Edition: everything in Data Access, plus asynchronous multi-site disaster recovery (AFM DR), native encryption, file audit logging and watch folder.
- Erasure Code Edition: Data Management features plus erasure coding across the internal disks of storage-rich servers.
- Developer Edition: Data Management features for test environments only, limited to 12 TB per cluster.
IBM also sells IBM Storage Scale System, the appliance formerly known as Elastic Storage System.
What Lustre is
The Lustre manual describes Lustre as a storage architecture for clusters whose file system “is supported on the Linux operating system” and offers a POSIX interface. It is best known for powering many of the largest HPC clusters, and many sites run it as a site-wide file system that serves several clusters. The software is licensed under GPL 2.0 for use with Linux.
The Lustre trademark is jointly owned by OpenSFS and EOFS, and community releases are coordinated in the OpenSFS Lustre Working Group. Its April 2026 release update shows DDN with the largest share of commits to Lustre 2.17, alongside HPE, AWS, ORNL and others. At the time of writing, the Lustre release page lists 2.17.0 as the current major release, 2.15.8 as the current LTS release and 2.18 as in development.
The manual also says Lustre does not suit “peer-to-peer” setups where every node is both client and server, citing the lack of data replication at the Lustre software level. File Level Redundancy (FLR), added in 2.11, lets individual files keep mirrors on several OSTs.
Architecture: metadata and data paths
Lustre splits roles across dedicated servers, as described in the Lustre wiki and the manual:
- The MGS stores configuration for all Lustre file systems in the cluster.
- The MDS handles namespace operations and serves metadata from one or more MDTs: file names, directories, permissions and file layout.
- OSS nodes serve file data stored as objects on OSTs. Clients send data I/O directly to the OSS nodes.
- DNE (Distributed Namespace Environment) adds MDTs. Remote directories place sub-trees on separate MDTs, and since 2.8 striped directories spread one directory over several MDTs.
Progressive File Layout (PFL, 2.10) gives a file several components with different stripe settings. Data on MDT (DoM, 2.11) keeps small files on the MDT so that small random I/O does not reach the OSTs.
Storage Scale uses a shared-disk model. A file system is a set of Network Shared Disks (NSDs) holding both data and metadata. Metadata files such as the root directory, allocation maps and inode file are allocated only on disks where metadata is allowed, and the administrator chooses those disks. Nodes reach disks directly over a SAN or through NSD servers over the network, a layout IBM says is often used in large clusters. Consistency is coordinated by two roles:
- One file system manager per file system handles configuration changes, disk space allocation, quotas and the tokens that keep data and metadata consistent (IBM docs).
- One metanode per open file applies that file’s metadata updates, while every node reads and writes file data directly (IBM docs).
Comparison table
| IBM Storage Scale (GPFS) | Lustre | |
|---|---|---|
| Licence | Commercial; current editions licensed per TiB of capacity | Open source, GPL 2.0 |
| Metadata design | Metadata on NSDs, on disks the admin allows for metadata; tokens from the file system manager; one metanode per open file | Dedicated MDS nodes serving MDTs; DNE spreads the namespace over multiple MDTs |
| Client OS support | Linux, AIX, Windows (Windows desktop editions as NSD clients only) | Linux |
| Tiering | Storage pools with placement and migration policies (ILM); tape with additional IBM software; AFM between clusters and NFS sources | OST pools, PFL, Data on MDT; HSM to external storage through a copytool |
| RDMA / InfiniBand | RDMA on Linux only, over InfiniBand, RoCE and Omni-Path; AIX and Windows use IP over InfiniBand | LNet supports InfiniBand and IP networks; RDMA through the matching Lustre network driver (LND) |
| Protocol access | NFS, SMB, S3 and HDFS through CES protocol nodes | Native Lustre client |
| Typical deployments | SAN-attached or NSD-server clusters, mixed-OS clusters, multi-cluster sharing over WAN | Large Linux HPC clusters and site-wide file systems serving many clusters |
Storage Scale OS and RDMA rows: IBM Storage Scale FAQ, questions 2.5, 3.2 and 6.3. Lustre networking and HSM: chapters 2 and 26 of the Lustre manual.
Operations and skills required
Running Lustre involves:
- High availability. The manual requires external HA software with resource fencing and power control that can switch off a failed server, so two nodes never access the same storage at once.
- Networking. LNet loads one LND per network type, such as
ko2iblndfor InfiniBand orksocklndfor TCP. Multi-Rail and routing between networks are configured on top. - Back-end storage. MDTs and OSTs sit on ldiskfs or ZFS, each with its own formatting, quota and repair tools.
- Layouts. OST pools are created with
lctl pool_new, andlfs setstripesets stripe counts, PFL components and the pool a file or directory uses. - Versions. Lustre packages are tested against specific enterprise Linux releases, so OS and kernel upgrades follow the Lustre release notes.
Running Storage Scale involves:
- One stack on every node. Every node runs the kernel extension and daemon. Manager and quorum roles are assigned to nodes, and the cluster manager picks a new file system manager if one fails.
- IBM tooling. Administration uses the
mm*commands (mmcrfs,mmlsmgr,mmchmgrand others), a GUI and a REST API. Tiering is driven by ILM policy rules for placement and migration. - Protocol nodes. CES moves IP addresses from a failed protocol node to a healthy one. IBM notes that SMB failover is not transparent, that all protocol nodes must run the same version, and that each needs a server licence.
Licensing and support models
IBM Storage Scale. According to the IBM FAQ (questions 14.2 and 14.3), the Data Access, Data Management and Erasure Code editions are licensed by capacity, per TiB (2^40 bytes). Licensed capacity is the provisioned capacity presented as NSDs, and data reduction does not change it. Per-drive licences apply when bought with IBM’s storage appliance. The older socket-licensed Standard and Advanced editions are sold only to existing licensees. Support comes from IBM. We don’t quote prices here.
Lustre. The software is GPL 2.0 and community releases are free to download. There is no default support contract: sites rely on their own staff and community resources such as the Lustre wiki and issue tracker, or buy support from a storage vendor or integrator.
Where BeeGFS fits
BeeGFS is the parallel file system alternative that often appears in the same comparison. The BeeGFS documentation describes separate management, metadata and storage services plus a Linux client, which puts it closer to Lustre than to Storage Scale. Key documented points:
- Metadata is distributed per directory, with each directory assigned to a randomly chosen metadata node.
- Clients and servers. The client is a Linux kernel module that needs no kernel patches. Servers are user-space daemons, and services can be combined on the same machine.
- Mirroring. Buddy mirroring mirrors file contents and metadata synchronously between pairs of targets.
- Networking. RDMA over InfiniBand, RoCE and Omni-Path, with TCP also available.
- Licensing. The client module is GPLv2, and other components fall under the BeeGFS License Agreement. In BeeGFS 8, enterprise features such as storage pools and quotas need an enterprise licence, and a community licence carries a capacity limit.
For head-to-head comparisons, see Lustre vs BeeGFS and Ceph vs BeeGFS.
Which to choose by workload
- Windows, AIX and Linux clients together, or NFS/SMB/S3 access to the same data: Storage Scale, which documents all three operating systems and integrated protocol nodes.
- Data shared or replicated between sites: Storage Scale AFM, with AFM DR in the Data Management Edition.
- Large Linux-only HPC scratch or a site-wide file system: Lustre, especially with Lustre skills in-house and a preference for GPL software.
- Tape or another tier behind the fast file system: both. Storage Scale uses ILM policies with IBM archive or backup software; Lustre uses HSM with a copytool.
- Mixed small and large files: both have mechanisms. Lustre offers Data on MDT and PFL; Storage Scale can store small files’ data in the inode.
- Linux HPC or AI clusters wanting separate metadata and storage services with RDMA: BeeGFS.
- Home directories and software trees with low concurrency: NFS is usually enough.
Mevasis designs parallel file systems and deploys BeeGFS as an official BeeGFS partner. If you are weighing these options for a new cluster, see our HPC storage solutions.
Frequently Asked Questions
Yes. IBM Storage Scale is the current name of the product previously called IBM Spectrum Scale and originally GPFS (General Parallel File System). IBM support pages still list Spectrum Scale and GPFS as alternate names.
The Lustre software is licensed under GPL 2.0 for use with Linux, and community releases are published through lustre.org. The costs are hardware, staff time and any commercial support contract you choose to buy.
The current Data Access, Data Management and Erasure Code editions are licensed by capacity, per TiB. A free Developer Edition exists for test environments, but it is limited to 12 TB per cluster and is not supported for production use.
IBM Storage Scale supports Linux, AIX and Windows nodes, although RDMA and protocol services are Linux-only. The Lustre manual describes Lustre as supported on the Linux operating system.
BeeGFS is a Linux parallel file system with separate management, metadata and storage services, which makes it structurally closer to Lustre. Its client module is GPLv2, the other components fall under the BeeGFS License Agreement, and features such as storage pools and quotas require an enterprise licence.
Related Articles
- BeeGFS Technical Guide: Architecture, Installation, and Best PracticesBeeGFS parallel filesystem architecture, four-component design (management, metadata, storage, client), installation steps, troubleshooting, …
- HPC Storage Technical Guide: Three-Tier Architecture, BeeGFS vs Lustre, and TroubleshootingHPC storage architecture technical guide: 3-tier design (NVMe scratch, parallel filesystem, capacity archive), BeeGFS vs Lustre decision …
- Lustre vs. BeeGFS: Choosing a Parallel Filesystem for HPCArchitecture, performance, management, and cost comparison of Lustre and BeeGFS parallel filesystems. A practical guide to HPC storage …
- Mevasis Is Now an Official BeeGFS PartnerMevasis has become an official partner for the BeeGFS parallel file system. What BeeGFS is, why it is chosen, and how licensing, deployment …
Let's Build Future Together.
