ABOUT LUSTRE

The parallel file system for HPC and AI. Any hardware. Any cloud. Any scale.

Lustre is the parallel file system powering the world’s fastest supercomputers and largest AI factories — and runs just as well on a handful of servers in a lab

WHAT LUSTRE IS

Lustre is an open-source, parallel file system designed for high performance, scalability and availability. It is purpose-built to provide a coherent, global, POSIX-compliant namespace for large-scale data processing, and is used by a majority of the world's largest supercomputers.

Thousands of organizations trust Lustre for three key reasons:

Open

Lustre is fully open source under the GPL, client and server alike. It is developed in the open by a community of national laboratories, vendors and cloud providers.

Proven

In production for more than twenty years. It runs on over half of the world's 100 fastest supercomputers, in the largest AI training clusters, and as a first-party managed service from every major cloud provider.

Sovereign

Your data, on hardware or a cloud you choose, on a file system no single company can withdraw or steer. With the freedom to support it and change it yourself, Lustre is a platform you own for as long as you run it.

HOW LUSTRE WORKS

Parallel by design

Lustre is built for massively parallel data processing, with thousands of clients reading and writing at the same time. Files are striped — or mirrored — across many servers, so every client reads and writes in parallel and shares the aggregate bandwidth of the whole system.

Many tiers, one namespace

Object storage servers can be grouped into pools by media; NVMe flash for performance, QLC flash for capacity, HDD for cold data.

Files can be placed on, migrated between or mirrored across them without changing their path or name. Data can also be transparently staged out to external storage through Lustre's Hierarchical Storage Management (HSM), and brought back automatically the next time it is read.

From a lab to exascale

Lustre scales from a single server in a lab to the largest systems on Earth: tens of thousands of clients, hundreds of petabytes, tens of terabytes a second, read and write, without any change to your applications.

Applications see a single, fully coherent, shared POSIX namespace, so existing software runs unchanged: no rewrites, no special libraries, no relaxed consistency to design around.

THE FEATURES THAT MATTER FOR HPC AND AI

Lustre has been adapted to demanding HPC workloads in production for over twenty years, and its feature set reflects that: broad, mature and built for the problems that appear at scale. The table below covers the capabilities that matter most for HPC and AI today.

    
      Performance
      
          Parallel I/O
          Supported: Files striped across servers; clients read and write in parallel.
        
          RDMA on every fabric
          Supported: InfiniBand, Ethernet/RoCE, Slingshot and more. Multi-rail LNet.
        
          GPUDirect Storage 2.15+
          Supported: Data straight into GPU memory.
        
          Persistent Client Cache 2.18
          Supported: Shared datasets cached read-only on each client’s NVMe: a local “tier 0”.
        
      Scale and data placement
      
          Distributed metadata (DNE)
          Supported: One namespace spread across many metadata servers.
        
          Heterogeneous media
          Supported: NVMe, TLC and QLC flash, HDD: pools in one namespace.
        
          Transparent migration
          Supported: Move files between pools. Path and name unchanged.
        
          Hierarchical Storage Management
          Supported: Stage out to S3, other file systems or tape; back on demand.
        
      Data protection and security
      
          Data mirroring (FLR)
          Supported: File-level replicas across servers or pools.
        
          Erasure coding 2.18
          Supported: Redundancy without full mirror overhead. Initially read-only, in 2.18.
        
          Encryption at rest 2.15+
          Supported: Client-side encryption with fscrypt, per directory or file.
        
          Transport encryption
          Supported: Kerberos or shared-key (SSK) secured traffic.
        
      Multi-tenancy and operations
      
          Multi-tenancy
          Supported: Secure sub-directory mounts with per-tenant ID and quota domains.
        
          Changelogs
          Supported: Audit trail of every namespace change. Drives replication tools.
        
          Kubernetes CSI
          Supported: Lustre volumes for containers.
        
          OpenTelemetry metrics
          Supported: OTLP collectors and Grafana dashboards.

Have a project in mind or need expert guidance? Send us a message and let’s explore how our team can help?

Get in touch →

TALK TO OUR EXPERTS