In this video, the creator details the construction and optimization of a petabyte-scale ZFS storage array for a home AI datacenter, achieving up to 17GB/s throughput using high-end hardware, careful cabling, and noise reduction techniques. The video covers hardware choices, performance testing, networking upgrades, and practical tips, while previewing future content on advanced ZFS configuration and management.
In this video, the creator showcases the fourth part of their high-performance AI home datacenter build, focusing on a petabyte-scale ZFS storage array capable of achieving up to 17–18 gigabytes per second throughput. The system is built around a Threadripper 5955WX CPU on a WRX80 motherboard, running Proxmox as the operating system. The storage headend features three Mellanox CX4 100-gigabit network cards and four SAS HBAs, each connected to a JBOD (Just a Bunch Of Disks) enclosure. The drives used are primarily Seagate Exos X14, X16, and X18 models, with a mix of 10TB and 14TB capacities, optimized for both performance and noise reduction by stacking them tightly in the rack.
The video delves into the technical details of maximizing storage performance, including the importance of multipath configurations for high throughput and the careful selection of PCIe cards and cabling. The creator emphasizes that using quality cables and hardware is crucial, as cutting corners often leads to performance bottlenecks and additional costs down the line. The ZFS array is currently set up with each JBOD as an independent VDEV in a large RAIDZ2 stripe, but the creator notes that mirrors or other configurations may be preferable for redundancy and rebuild times. Performance testing is demonstrated with ZFS ARC caches disabled to prove that the high speeds are coming directly from the drives.
A significant portion of the video is dedicated to hardware modifications and optimizations, such as removing air dampers from JBOD power supplies to reduce noise and ensuring proper airflow. The creator also discusses the differences between NetApp DS2246 and DS224C enclosures, highlighting the importance of SAS backplane versions for achieving higher transfer rates. Practical tips are shared for cabling and connecting HBAs to JBODs, including the correct use of ports and the impact of different cabling methods on performance and redundancy.
Networking and infrastructure improvements are also covered, with upgrades to the structured media center that serves as the demarcation point between the house and garage networks. The setup now includes UniFi switches, a dedicated 10-gigabit switch, improved fiber management, and an APC UPS for extended runtime during power outages. The creator stresses the importance of aligning the storage system’s performance with the intended workload, particularly for handling large AI tensor files, and recommends specific networking hardware for reliability and efficiency.
Finally, the video wraps up with a look at noise reduction strategies, further hardware recommendations, and a preview of upcoming content focused on ZFS configuration and scripting for multi-JBOD setups. The creator teases a new ZFS profiling tool and plans for a massive SAS SSD metadata array to complement the high-capacity spinning disks. Viewers are encouraged to share their own experiences with high-performance storage arrays and to stay tuned for the next installment, which will dive deeper into operating system and ZFS setup details.