# PEAK:AIO’s Lattice open-source pNFS metadata server speeds AI data feeds > Source: > Published: 2026-06-04 16:06:47+00:00 # PEAK:AIO’s Lattice open-source pNFS metadata server speeds AI data feeds [PEAK:AIO](https://www.blocksandfiles.com/ai-ml/2025/10/06/peakaio-banks-68m-to-scale-out-ai-storage-platform/1604254) and Los Alamos national Laboratory (LANL) have developed a dynamically scalable, open-source metadata server for parallel NFS, permitting up to 1,000 metadata servers and enabling higher-speed file data transmission to GPU sercers NVS v4.1 enabled -[parallel NFS](https://www.blocksandfiles.com/glossary/2022/05/05/pnfs/1586765) (pNFS) to speed file delivery to clients. This requires a metadata server (MDS) to provide file layouts to client systems, which use them to perform direct, parallel reads/writes to the data servers. The layouts describe where and how a file's data is striped across data servers, including locations, striping parameters, and access tokens. Up until now these metadata servers have been proprietary and not very scalable. PEAK:AIO says AI workloads, including training large models, running inference at scale and serving agentic AI applications, require ultrafast, parallel access to massive datasets continuously and reliably. GPU compute has scaled dramatically, but the storage layer feeding it, and particularly the metadata architecture coordinating it, has not kept pace. Average GPU utilization across 23,000 production clusters is 5 percent, according to [Cast AI](https://cast.ai/reports/state-of-kubernetes-optimization/), not because the hardware is inadequate, but because the software feeding it cannot keep up. AI workload performance is constrained because there is a metadata bottleneck in [parallel storage systems](https://www.blocksandfiles.com/ai-ml/2025/11/26/parallel-file-systems-explained-metadata-striping-and-throughput/1707837 ). PEAK:AIO’s PNFS-Lattice software fixes this problem. (We discussed its initial incarnation [here](https://www.blocksandfiles.com/ai-ml/2025/11/12/peakaios-scale-out-flash-pnfs-metadata-server/1717703).) Gary Grider, LANL HPC Division Leader, said: "PNFS-Lattice is unique in that it is an open-source, user-space, scalable PNFS metadata server, from the ground up, by leveraging the concept of separating the PNFS metadata service from the Metadata Store (catalog).” “Since the service is separate from the persistent metadata and it runs in user space, it is well poised to be an ephemeral service that could be resized on the fly. Further, since it is open-source and user space, it lowers the bar for community participation, encouraging more innovation driven by AI, HPC, and other community needs." Lattice separates the metadata control plane into four distinct layers: Protocol State Plane, Lattice Core, MD Catalog Authority and Data Server Control Plane. This architecture makes metadata services truly elastic for the first time, allowing them to spin up dynamically on commodity hardware whenever and wherever needed, from a single server to more than 1,000 metadata servers. Mark Klarzynski, CSO and Co-Founder of PEAK:AIO, said: “That separation unlocks intelligent scale in a way traditional storage architectures were never designed to support. Metadata and data services can now become distributed, elastic participants that scale, fail over and adapt around the workload, rather than remaining fixed appliances or static MDS pairs. This is a fundamental step forward for pNFS and parallel file system design for ultra-high-performance storage, allowing metadata to move beyond the limitations that have constrained scale-out storage for decades.” The software is open-source, community-supported, and launched under the Linux Foundation. Performance testing conducted during the PEAK:AIO and LANL collaboration demonstrated gains from 70 GB/s to 400 GB/s. On existing production hardware at LANL, standard Linux NFS configurations delivered between 3 GB/s and 7 GB/s throughput, while the pNFS Lattice architecture achieved 40 GB/s on identical servers. Additional testing conducted with a Tier 1 technical university demonstrated metadata-heavy workload improvements exceeding 300 percent compared with conventional approaches. In standard metadata benchmarks such as MDtest, early testing has demonstrated up to a 10x improvement over standard Linux KNFSD, while Lattice’s advanced features have delivered more than 300 percent improvement in traditionally difficult metadata-heavy workloads. PEAK:AIO says that, “when combined with its elastic, ephemeral metadata scaling model, where metadata services can be added dynamically as demand grows, Lattice moves beyond the limits of conventional high-performance data designs and creates a new foundation for metadata performance, resilience and scale in open pNFS and parallel file system design.” Roger Cummings, President and CEO of PEAK:AIO, said: “AI infrastructure markets are approaching an inflection point where scaling compute alone no longer delivers any meaningful efficiency gains. Our collaboration with Los Alamos National Laboratory was built around the idea that if AI infrastructure is to scale efficiently, metadata must become elastic, distributed and open. Lattice represents that transition, and we’re excited to build it with the Linux and HPC communities beside us.” PEAK:AIO will provide PEAK:AIO pNFS, a commercially supported superset of Lattice, for organizations that want enterprise SLAs and a full featureset without managing the open-source stack directly. This model mirrors the relationship between Lustre and its commercial distributions, while maintaining a fully open standards-based foundation.