The parallel filesystem the world's largest GPU clusters trust — deployed as a two-node cluster in your cloud account, with object storage as the capacity layer. Lustre, NFS, and SMB serve one namespace from one pool, and one terraform apply stands the whole thing up.
Both nodes run active-active — metadata and object services float between them with the virtual IPs, protected by a witness against split-brain. The data lives in object storage buckets behind ZFS, with NVMe for metadata and SSD for the log — so capacity grows at bucket prices while clients see a normal high-performance filesystem.
Unified Lustre architecture on Google Cloud — the same shape deploys on AWS and Azure.
Lustre for GPU training, NFS for tools and notebooks, SMB for everyone else — the same files, the same pool, no copies between silos.
Data lives in S3, GCS, or Blob buckets behind the filesystem — capacity grows at bucket prices, durability comes from the cloud, and nothing is staged or copied in.
Every service can run on either node; roles and virtual IPs migrate on failure with a witness guarding against split-brain. Clients reconnect and training continues.
Stock Lustre and NFS clients — GPU instances, managed Kubernetes, existing HPC images. Nothing proprietary to install or maintain on the compute side.
Every resource deploys into your own project and VPC. No vendor infrastructure, no call-home, no external endpoints — you keep full IAM, network, and data ownership.
A complete Lustre cluster — metadata, object services, and NAS — on a two-node footprint. The operational surface of a filer, the throughput profile of a parallel filesystem.
Google and Microsoft each ship a Lustre CSI driver with their managed Kubernetes service — and both of them mount this filesystem. The same PersistentVolumeClaim, the same ReadWriteMany volume, the same YAML your workloads already have. Each driver supplies the Lustre client on the nodes, so pods mount a shared filesystem with nothing to build and nothing of ours running in your cluster.
The workload layer is identical on both clouds — same driver, same claim, same mount. Only the way each cloud floats an address differs. Addresses shown are from live clusters.
Open Lustre Community Edition is the whole picture — two storage nodes, active-active HA, every protocol — deployed into your own account and free to run. There is no managed service to buy, no marketplace subscription, and nothing metered. Bring it up for a semester, a grant milestone, or an afternoon of experiments, and tear it down when you are finished. Talk to us for sizing, support, or additional HA pairs.
Self-serve with terraform from your cloud marketplace — ready for your most demanding training runs.
