LeilFS Adds NVMe Flash Support – Independently Benchmarked on WD OpenFlex Data24.

Not Native Yet, Already Fast 

Basic, production-ready NVMe (flash) support comes to LeilFS — our open-source, HDD-native distributed file system

We added production-ready NVMe flash support to the community edition of LeilFS, our open-source distributed POSIX file system, and WD independently benchmarked it on the OpenFlex Data24 NVMe-oF platform – sustaining up to 40 GB/s reads 36 GB/s writes per OpenFlex Data24 enclosure.

LeilFS – the open-source, community edition of our HDD-native distributed file system – now supports NVMe flash storage. We want to be precise about what that means, because precision is the whole point of a storage vendor you can trust. 

So let us start with the honest headline: this is basic production-ready NVMe (flash) support, and we are not NVMe-native yet. For hard drives, including SMR, we describe LeilFS and LeilOS as HDD-native, and we mean it literally: the software runs the media at close to its full capability. For flash we have not earned that word yet. What we have earned is basic, production-ready NVMe support. You can put real flash workloads on LeilFS today. 

The reason we are comfortable saying production-ready, and not merely functional, comes down to the number that tends to matter most in practice: the metric that matters for NVMe storage in practice: sequential write throughput, measured in GB/s. Write throughput is one of the important metrics because unlike reads it’s typically negatively impacted by extra data redundancy (the extra data consumes writes bandwidth) so it’s crucial to keep writes at appropriate level. 

Independently benchmarked by WD on OpenFlex Data24 (NVMe-oF), and built for a fair fight 

WD ran the benchmarks themselves, in their own lab, on their own hardware, using the standard fio tool, and published the results in their solution brief for LeilFS on OpenFlex Data24. The setup was a single 24-bay NVMe-oF enclosure filled with 24 standard NVMe SSDs, shared over a 100GbE fabric to six storage nodes and six clients. Reads are solid. Writes are where the story gets interesting. 

The enclosure hands out raw NVMe namespaces with no built-in RAID, so every software stack on top of it builds and pays for its own data protection. That makes for a genuinely fair contest, as long as you compare like with like. WDand its partners have published results for several other stacks on this exact platform, which lets us line LeilFS up against the field. Because these are other companies’ products, we will describe them by category rather than by name. 

Read those rows against the protection each stack provides. Against the most widely deployed open-source object storage, LeilFS writes more than three times faster at the same erasure-coding protection. Against an established commercial parallel filesystem running single-parity RAID, LeilFS keeps pace while carrying double parity, the more expensive and more resilient choice, which means at equal protection we would expect to move ahead. And even a bare NVMe-oF initiator that adds no filesystem and no protection at all does not leave LeilFS behind on writes. For a first flash release, that is a strong place to start. 

Why LeilFS Delivers Fast NVMe Write Throughput from Day One: SMR Sequential-Write Optimization

None of this is luck, and it is not a flash-specific rewrite either. It comes from where LeilFS grew up. 

LeilFS is HDD-native, and the hardest thing to do well on a modern hard drive, especially a shingled SMR drive, is write. SMR forces you to write in large, sequential, zone-aligned streams, so we spent years building a write path that organizes data into big sequential runs and lays it down in the order the media actually wants. 

Flash rewards almost exactly the same discipline. An SSD dislikes small scattered writes, which trigger garbage collection and write amplification, and it loves large sequential ones that map cleanly onto its erase blocks. The sequential-write algorithms we built for shingled HDD carried over to NVMe with very little translation, and they are the reason our writes were competitive from the very first flash build. The optimization we did for the slowest media turned out to be exactly right for the fastest. 

What’s Next: Host-Managed Over-Provisioning to Lower Flash TCO for QLC SSDs

Basic support is a foundation, not a destination, and the work ahead is not really about squeezing out more gigabytes per second. It is about the economics of the media itself. It echoes an argument we made for hard drives in “A Failing Head Is Not a Failed Drive.” A single failing component should not condemn an entire drive. 

Flash has the same waste in a different form. A modern QLC SSD is managed as a black box. Its firmware quietly burns through spare blocks to hide failing dies, and when that reserve runs out it retires the whole drive, often while most of its media is still perfectly good. At datacenter scale, that is a great deal of usable flash written off on the firmware’s schedule rather than yours. 

We think that decision belongs to the filesystem, not the firmware. Instead of failing, a drive should gracefully shrink. The host reads the drive’s real health telemetry, evacuates data from degrading regions ahead of time, and reduces the drive’s usable capacity to reclaim performance and extend its life, while cluster-level erasure coding keeps everything durable underneath. Wear-out stops being a sudden write-off and becomes a managed, predictable expense. That is a real lever on flash total cost of ownership (TCO). 

Meet Leil at SNIA SDC 2026: Host-Managed Over-Provisioning for the QLC Era

We are presenting exactly this at the SNIA Storage Developer Conference in September, in a talk on host-managed over-provisioning for the QLC era. If you run flash at scale, or you are simply tired of retiring drives that still have most of their life left, come find the team. Details and how to meet us are here: Meet us at SNIA Developer Conference SDC 2026. 

Basic NVMe support is live in the open-source LeilFS distributed file system today, its writes already hold their own, and the most interesting part is still ahead.

We would love to hear about your workload, and especially about your drive-retirement numbers. Those are the ones we are building to fix. 

Author: David Gerstein
David Gerstein is the CTO of Leil, an Estonia-based storage infrastructure company building HDD-native software for the exabyte era. Leil is in active co-engineering partnership with Western Digital, validated across the WD, Seagate, and Toshiba drive ecosystems.