Here’s an interesting concept: ZettaLane has built MayaNAS, a Lustre parallel filesystem supporting NFS, pNFS and SMB, which runs atop cloud object storage.
MayaNAS can stream AI training at GPU‑level speeds, deliver tens‑of‑gigabytes‑per‑second checkpoint performance, and match object‑bucket cost profiles. Validated via MLPerf Storage benchmarks, it hit 32.42 GB/s Llama 3 70B checkpoint write bandwidth across two clients. For RetinaNet workloads, one MayaScale 200 Gb/s client fed 72 B200 GPUs.
Headquartered in Santa Clara, California, ZettaLane was founded in June 2018 by CTO Supramani (Sam) Sammandam, a seasoned storage engineer with over 25 years of systems‑software experience at Lucent Technologies, IBM, HP and Dell. The firm operates entirely without venture‑capital backing.
MayaNAS combines Lustre and OpenZFS to reshape storage economics. Each OST (Object Storage Target) functions as an OpenZFS dataset whose vdevs point to regional Standard‑class Google Cloud Storage buckets. Bulk data reads and writes go concurrently to object storage through its objbacker layer, while a compact NVMe‑based special vdev stores only pool metadata and small blocks. No local SSD or ephemeral scratch tier sits within the data path.
Supramani (Sam) Sammandam commented: “Within MLPerf Storage v3.0 results, ZettaLane’s MayaNAS was the sole entry running standard Lustre directly on cloud object storage — using regional GCS in the data path on generic cloud VMs, no local SSD, operating inside the customer’s own cloud account.
“Verified by MLCommons, it occupies a different category from flash‑based submissions such as Everpure. It delivers parallel‑filesystem throughput paired with object‑storage economics, rather than focusing purely on raw GiB/s numbers. It executes full pipelines without staging data onto NVMe, including POSIX‑dependent KV caches that S3‑native platforms cannot support.”
We raised a comparison point: does MayaNAS resemble Nasuni, which builds file services upon object‑storage backends?
Sammandam responded: “At a high level there is similarity: both place a filesystem on object storage. Many competitors build brand‑new filesystems and expend great effort proving POSIX compliance. We took an alternate path, building on field‑proven OpenZFS with its native split between metadata held on small NVMe devices and bulk data residing on object storage. We shared this work at the 2025 OpenZFS Developer Summit and presented our Lustre design at LUG 2026.
“Nasuni follows a cache‑first model: edge filers serve traffic from local caches while authoritative copies remain in object storage, so performance depends heavily on working‑set cache fit. MayaNAS reads and writes directly to object with no intermediate local data cache. In MLPerf testing, cold object reads satisfied GPU‑scale workloads, confirmed via native Google Cloud Monitoring metrics.
“The more fundamental distinction lies in its nature: MayaNAS is a standard Lustre and pNFS parallel filesystem, with NFS/SMB capabilities, working with unmodified in‑kernel Linux clients. It is purpose‑built for GPU‑scale parallel throughput, making it well‑suited for benchmarks like MLPerf. Our cloud‑native pNFS implementation leverages open‑source MDS technology from LANL. Nasuni targets distributed enterprise file use‑cases; MayaNAS covers enterprise NAS while extending directly into HPC and AI parallel workloads on object storage.
“That captures the core idea. Object storage is clearly the forward direction for AI infrastructure, yet adoption generally forces application rewrites for S3 compatibility. MayaNAS resolves this: standard Lustre or pNFS running over object storage, granting object‑level cost benefits with zero code changes.
“There are further components under this architecture: OpenZFS‑on‑object internals, pNFS together with LANL MDS integration, MayaScale NVMe‑over‑TCP block storage, and branchable Postgres for agentic AI, all built upon this common foundation.”
Bootnote
After publication, Sammandam provided additional updates: “ZettaLane began as an LLC back in 2018. I managed it part‑time alongside other career roles, including a technical staff position within Dell ISG’s Office of the CTO, where I observed VAST and WEKA gaining traction in AI infrastructure. I returned full‑time in 2026 to advance parallel‑storage development, and MLCommons’ MLPerf validation gave confidence to bring the product to public view. These benchmarks exercise datasets larger than 20 TB alongside millions of files, with results peer‑reviewed by other submitters including major corporations.
“Funding remains fully bootstrapped and self‑financed with no external investment, as previously noted.
“To be transparent, we do not yet have prominent production customer deployments. That represents our next key milestone, and these MLPerf results serve as our primary proof‑of‑value for market outreach. We hope to secure collaboration from one of the leading object‑storage vendors.”
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Người liên hệ: Ms. Sandy Yang
Tel: 13426366826
Vietnamese
