At the 66th IT Press Tour in San Jose last week, VMblog sat down with Peter Jones (CEO) and Andreas Dilger (Principal Architect) from The Lustre Collective, a newly formed company focused entirely on advancing Lustre, one of the most battle-tested parallel filesystems in existence.
Here’s what we learned about why this matters, and where Lustre is heading.
A Filesystem With Real Pedigree
Let’s start with some context. Lustre has been around for 25 years. It wasn’t built as a startup’s moonshot idea – it came out of U.S. Department of Energy labs with the explicit goal of breaking IBM’s monopoly on high-performance filesystems at the time. That origin story matters because it shaped what Lustre became: a no-nonsense storage system built to handle massive, pathological workloads that would make other filesystems throw up their hands.
The proof is in the numbers. Lustre consistently powers between 60% and 70% of the top 100 HPC systems globally. More recently, it’s become the filesystem of choice for the world’s largest AI systems, including NVIDIA’s DGX SuperPOD, xAI’s Colossus supercomputer, and the Frontier and El Capitan exascale systems. Even the hyperscalers have embraced it: AWS, Google, Microsoft, and others all offer Lustre-based first-party services.
What’s remarkable here is that Lustre did this while remaining open source under a GPL version 2 license. That permissive approach meant vendors couldn’t hide behind proprietary wrappers. Fujitsu, Cray, and others used Lustre as their foundation, but they also invested back into the ecosystem. And companies like Hudson River Trading, a high-frequency trading firm with seasoned HPC storage admins, have consistently chosen Lustre even when evaluating every competing option. As they put it during a presentation at the Lustre User Group: “despite marketing claims that might come from contemporary technologies being heavily marketed, we find that we can always return to Lustre because it consistently just works.”
Why Start a New Company Around a 25-Year-Old Project?
Here’s the interesting part: The Lustre Collective is unconventional for a startup. Usually, you start with an idea and convince people it’s worth building. TLC started with a product people already know, trust, and use – but it’s free. That creates a different business problem.
The core issue is that Lustre needed an independent organization with a clear mission. For years, development responsibility bounced around: Cluster File Systems (the original company) was acquired by Sun, which was acquired by Oracle, which had no interest in HPC. Then a group of Lustre developers spun out to form Whamcloud, which got acquired by Intel. Intel used Lustre as part of its broader HPC ambitions, but that created conflicts. Some companies were hesitant to work with Intel’s Lustre team because of competitive concerns around other Intel technologies.
Then DDN acquired the Whamcloud assets. DDN is a legitimate storage vendor, but as a vendor, DDN’s incentives naturally focus on its own portfolio and shareholder returns. Lustre became one component in a broader strategy rather than the focal point.
The Lustre Collective changes that equation. It’s not seeking venture capital, and the engineering team retains control. The goal is to run it more like a foundation – an ongoing concern that exists to serve the community and advance the technology.
“We’re not looking to sell or do an IPO,” Jones explained. “We’re looking for this to be ongoing, more akin to a foundation.”
How Does a Free Product Make Money?
This is the practical question, and TLC has a straightforward answer: services. Lustre itself remains free, but the company offers consulting, training, support, performance tuning, feature development, and deployment services. They’ve already fielded inquiries and signed actual contracts in several of these areas.
The advantage? Lustre is well-established enough that people know what they’re getting. “Because we’re well established in the market, people know Lustre and they trust that we’re able to deliver,” Dilger noted. That means TLC doesn’t have to advertise, potential customers are already coming to them.
The team’s credentials help. Jones has spent decades managing Lustre development teams across various companies. Dilger is essentially a living institution in the Lustre community. He’s been involved since 1999, and there’s a running joke that he once accounted for five of the top ten all-time contributors on the Lustre mailing list under different email addresses. Colin Faber, the third founder, has been involved from the earliest days and has bootstrapped multiple well-known Lustre storage appliances.
Modern Lustre: Faster Than You’d Think
Andreas Dilger walked through the technical specifications, and the numbers are legitimately impressive. Lustre can achieve 10+ terabytes per second in read/write performance, with 100+ million IOPS. It scales to over 700 petabytes of capacity with more than 100 billion files. It can support 20,000+ simultaneous clients and 100,000+ GPUs connected to a single filesystem.
But here’s what’s interesting: Lustre was originally designed for hard drives, and some people assume that makes it dated. Dilger pushed back on that directly. Modern Lustre is actually very efficient with QLC (quad-level cell) NAND flash because it leverages large aggregate I/Os. “We don’t have to do any wear leveling or anything because we actually utilize the storage very efficiently,” he said.
Beyond raw speed, Lustre offers practical flexibility that newer, flashy filesystems sometimes overlook. It’s POSIX-compliant even when distributed across tens of thousands of nodes. You can re-export it via NFS, SMB, or S3 protocols. It supports multiple storage tiers-flash, disk, even NVMe directly attached to compute nodes-without forcing you to explicitly stage data between tiers. Clients can access data directly from whichever storage tier it lives on.
This matters for machine learning workloads, where you might want to cache training data locally on GPU nodes while keeping the full dataset on slower, cheaper storage. Lustre handles that transparently.
The Cloud Complication: Multi-Tenancy at Scale
As Lustre has moved into cloud environments, new challenges emerged. Cloud vendors buy massive GPU clusters with thousands of GPUs and want to lease them out in an on-demand fashion. They need to subdivide that single, aggregate storage without physically partitioning the hardware.
Lustre addressed this by virtualizing both the storage and its management. Different users or groups can be isolated from each other, not just for security, but for capacity limitations and performance guarantees. The flexibility here is substantial. You can create virtual filesystems with different performance characteristics, apply fine-grained access controls, and isolate even sensitive workloads within a single cloud deployment.
This is something that cloud vendors’ own first-party Lustre services can’t easily provide. AWS, Google, and Microsoft focus on their own infrastructure constraints. But customers within those clouds increasingly want features like multi-department isolation or the ability to serve different data sensitivity levels from the same cluster. That’s where TLC’s expertise becomes valuable.
What’s Coming in 2.18 and Beyond
TLC released a roadmap showing several significant features in development:
- Erasure Coded Files will provide data redundancy at the filesystem level rather than requiring hardware redundancy. This means a Lustre cluster can withstand whole server or rack failures transparently, without applications even noticing. This is particularly important for the cloud deployments where you might not have the luxury of hardware redundancy.
- Trash-Can/Undelete functionality lets users recover deleted files without administrator intervention. This sounds simple but addresses a real problem. Lustre can delete files extremely fast, and a rogue script with a wildcard in the wrong place can wipe data in seconds. A recoverable trash bin is a practical safeguard.
- Improved Metadata Redundancy and Metadata Writeback Cache are longer-term goals. The writeback cache is particularly interesting because it allows metadata operations like creates and permission changes to be completely asynchronous. Early testing showed nearly 100x speedups for workloads hammering a single directory.
- Client-Side Data Compression, Fault-Tolerant Management Server, and GPU Peer-to-Peer RDMA are among other features under development.
The roadmap also shows where contributors come from: DDN, ORNL, HPE, AWS, and others. TLC will be taking a larger role in feature development going forward, but they’re explicitly positioning themselves as supplementing, not displacing, existing contributions.
Lustre vs. Everything Else
We asked Dilger directly about PNFS (Parallel NFS), which some have suggested as a modern alternative to Lustre. His answer illuminated why Lustre has endured:
PNFS is fundamentally one-to-one at its core-a metadata server directs clients to specific data servers, but you can’t actually achieve true parallel I/O on a single file. You also can’t have multiple readers and writers on the same file simultaneously without explicit serialization. Lustre, by contrast, supports parallel access even within a single file. Files can be split across multiple storage targets with full cache coherency, and different parts of a file can live on different storage types, some on flash, some on disk, all transparent to applications.
This flexibility extends to sophisticated data migration. Writes might initially hit high-performance servers, but Lustre can transparently migrate data to slower, cheaper storage in the background based on file age and capacity. If there are cache hits, reads come from the high-performance tier. If not, they transparently read from slower storage.
Dilger noted that this striping can go up to 2,000 servers wide. The key to scalability is that clients learn the layout of a file when they open it and don’t need to query the metadata server for every I/O operation. Compare that to systems like Ceph, which chunk everything into 32-megabyte pieces and require more metadata lookups.
The Bigger Picture
There’s a broader theme running through TLC’s pitch. The company sees itself as growing the pie, not fighting for a slice. Lustre’s dominance in HPC and AI is well-established. But there’s still plenty of room for improvement and new use cases.
Not everything in the computing world is AI, Dilger emphasized. Universities, national laboratories, and various research institutions still need storage for traditional HPC, weather modeling, molecular dynamics, genomics-workloads that have nothing to do with machine learning but need the same high-performance, reliable storage that Lustre provides.
One final note: There’s a rumor circulating that DDN sold off its Lustre development team and the technology is effectively dead. That’s not the case. TLC is open to working with DDN, which continues to maintain a substantial team. The two organizations have different focuses – DDN on its storage appliance business, TLC on ensuring the core Lustre technology remains current and community-driven.
What This Means
The formation of The Lustre Collective signals confidence that a mature, open-source filesystem can remain competitive and relevant through independent stewardship. It’s an unusual approach in an industry that typically gravitates toward proprietary solutions or cloud-vendor lock-in, but Lustre’s track record suggests there’s something durable here.
For data center administrators and engineers evaluating storage for HPC or AI workloads, TLC’s existence means there’s now an organization with explicit accountability to the Lustre community and no conflicting commercial priorities pulling in other directions. Whether you’re deploying Lustre in a cloud environment, on-premises, or across hybrid infrastructure, that clarity of purpose could matter.
The company is young, just two months old at the time of this briefing, but the people behind it have decades of collective experience. They’re not starting from scratch with an unproven idea. They’re stewarding something that’s already proven itself across the world’s most demanding computing environments.
##
