Director of Applied Systems at Hammerspace, live from Supercomputing 2025. Together, they explore the performance engineering breakthroughs that enabled Hammerspace and Samsung to deliver a historic IO500 10 Node Production result using only standard Linux, the upstream NFSv4.2 client, and off-the-shelf NVMe hardware.
This episode breaks down how the Hammerspace Data Platform delivered more than a 33 percent gain over earlier submissions, doubled overall bandwidth, and achieved an unprecedented 809 percent improvement in the IO Hard Read test using Samsung PM1753 Gen 5 NVMe SSDs. Jonathan explains the Linux kernel innovations, metadata advancements, IO path optimization, parallel file system breakthroughs, and multi-instance file placement strategies that allowed Hammerspace to reach genuine HPC-class performance without proprietary clients or custom networking.
Listeners get a detailed walkthrough of the architectural differences between Research and Production IO500 submissions, the impact of metadata redundancy, the performance benefits of NFSd direct and NFS direct, the role of ZFS locking improvements, and how upstream Linux contributions directly advanced the state of HPC and AI data infrastructure. Jonathan also highlights the evolution of MLPerf benchmarking, the benefits of tier zero storage, and how Hammerspace performance engineering is unlocking new levels of efficiency and scalability for AI training, scientific workloads, and large-scale analytics.
This episode is essential for AI architects, HPC engineers, kernel developers, data scientists, and infrastructure leaders building the next generation of high-performance data platforms.
