Hammerspace says IO500 results show its standard Linux plus NFS system software achieves HPC-class performance without proprietary parallel file system complexity.
The IO500 benchmark ranks the performance of storage systems supplying data to supercomputers and other HPC (high-performance computing) systems, with its 10-Node Production result limiting them to serving just 10 clients. Hammerspace says that, for the first time, a fully standards-based architecture — standard Linux, the upstream NFSv4.2 client, and commodity NVMe flash — has delivered a 10-node Production fully reproducible IO500 result traditionally achievable only by proprietary parallel filesystems. There are 33 systems listed in the IO500’s SC25 10-Node Production results, and this was the fastest NFS result ever recorded, putting Hammerspoace at number 18 in the rankings.
Trond Myklebust, Hammerspace CTO and Linux NFS client kernel maintainer, said: “This IO500 result rewrites long-standing assumptions about what standards-based Linux and NFS are capable of. Achieving a leading 10-Node Production score with the Hammerspace parallel global file system using upstream Linux, pNFS and NVMe hardware demonstrates that HPC-class performance no longer requires proprietary clients or specialized file systems. This achievement is a significant moment for the Linux performance community.”
Hammerspace said its Data Platform software system was running in ten DSX nodes and used Samsung PM1753 NVMe SSDs. It “unified the storage across all DSX nodes into a single shared filesystem, which clients mounted with parallel NFS (pNFS) v4.2 using flexible files layouts.”
There was an overall score of 85.23 with 74.66 GBps bandwidth and 97.29 on the calculated IOPS kIOPS measure, putting it “alongside conventional HPC file system architectures like Lustre, DAOS, and WEKA – but without the requirement for specialized networking, proprietary clients, and specialized hardware.”
It said the 10 client machines connected to the storage system used a standard Linux NFS client; no oroprietary front-end client software was needed.
Hammerspace had previously made an IO500 submission at ISC25, which fared less well than this latest submission with Samsung;
It said: “Recent Linux kernel enhancements engineered and contributed upstream by Hammerspace were responsible for the improved scores, as were major improvements in Hammerspace code.”
Submission system configuration differences, such as the Samsung SSDs, also made a difference. The decline in the kIOPS number was due to the use of redundant Anvil metadata servers in the Production configuration at Samsung. There was only one such server in the prior, research-focussed, submission.
Hammerspace asserts that, “with the right architecture, standard protocols can achieve the low latency, massive parallelism, and global scalability required for AI training, RAG pipelines, agentic workflows, and classical HPC simulations.” The standard protocols being NFS v4.1 and pNFS.
This means it can equal HPC-class performance from parallel file systems such as DDN’s EXAScaler (Lustre), other Lustre systems, WEKA, Spectrum Scale and VAST Data, and do so without lock-in or expensive hardware and complex software. This positions it well for HPC-class AI training and inferencing workloads.
Read more in a Hammerspace blog.
