<!– HERO IMAGE: Photo of the DL380 G9 in your rack, front panel with drives visible –>

Show Image

Compute at home, home lab, what ever you wish to call it. I have always had something that allows me to tinker and learn. It has changed from time to time. I went through early on the phase of enterprise hardware running everything, before realising the power consumption of that early 2000’era gear, through to the cloud (unsure if I can call this a home lab), consumer gear and now back to a somewhat modern and efficient enterprise server.

This post is about where I have landed, why I landed here and why I think decommissioned enterprise hardware is the best low cost entry path for self-hosting that is leaps and bounds better than a domestic NAS.

Before I go any further, a disclaimer. My home lab is not a lab in the traditional sense. I am not spinning up and tearing down environments for certifications or learning new hypervisors. Yes I still do this, but my home lab is about running real workloads for my household in a reliable manner. It doesn’t power down and if it does I have a house of unhappy people.

Frigate NVR watching 7 cameras. Home Assistant automating my house. Immich managing our family photos. Plex serving media, and OPNsense providing routing for my PPoE internet connection and DHCP.

This blog you are reading right now, served from WordPress (Docker Compose). I class all of these as production workloads for my family and they need to be reliable.

I am all in.

That said, the home lab does serve a testing purpose. I do try new things to bring continue my learning journey, for example I stood up an Openclaw instance (the internet would make you believe you need a Mac Mini), try Docker containers / VM’s along with general learning. But my primary function is reliability and the ability to run services.

Why Enterprise Hardware?
The cloud is great, but it is expensive and can be limited. There are 3 laws of IOT and one of them being around latent sensitivity. Over the years I have learnt a few things

1. Latency Matters : IP Cameras is a prime example. Whilst having free access to AWS and Azure, running IP cameras in the cloud just doesn’t work without really fast internet pipes. My 50mb upload isn’t enough. These need to be ran locally, each camera is RTSP’ing close to 10megabits. I have 7 of these in my house
2. CPU Grunt : Low power devices are nice, but it is a balance of quality of life. My rack of Raspberry PI 5’s are simply no match for heavy workloads (Immich, HomeAssistant, Frigate) that my family uses and 2 x low power semi modern Xeon’s make short work of them
3. Cost To Buy : The cost of a modern NAS can be quite expensive. I paid $250 for a DL380 G9 (no rails) with 64GB of RAM. Simply put the compute is basically free. Storage is a universal given but compute is in effect given away. I will talk about Cost To Run below
4. Reliability : Having lived in the HP & Cisco eco-system for 25+ years broadly speaking enterprise hardware is solid gear. If you are not a enteprise hardware guy. The DL380 is like the Toyota Camry, they are so common parts are not an issue. ECC Ram multiple PSU’s etc . Running OPNsense I need relaibility
   
   

The thesis is simple. Decommissioned data centre hardware is really cheap, purpose built to run 24/7 and in my opinion is a far better foundation than any consumer NAS appliance you can buy.

This is gear that comes out of data centres as organisations refresh their fleets. It was designed for continuous operation. It has features that consumer hardware simply does not offer. ECC memory protects against bit flips. Redundant power supplies keep you running during a PSU failure. Hot swap drive bays let you replace failed drives without downtime. Out of band management via iLO (Integrated Lights-Out) gives you remote console access even when the OS is unresponsive.

Having ran fleets of these in the past either for myself or past employers they just kind of work.

The hardware is worth almost nothing. A server that would have cost tens of thousands of dollars when new can be picked up on eBay for a few hundred dollars and that is what I did.

Why You Should Not USe Enterprise Hardware?
I just told you why I like Enterprise Hardware (and I really do, and will conitinue to buy).

My approach on Enterprise Hardware may not be for you, it is for me, but it may not be for you. To be balanced here are a few points on why and to be honest I think if any of these are an issue, then a consumer product may be for you

1. Power Consumption : No sugar coating this, there is a cost to running any sort of gear like this. It has got a lot better, especially if you are looking at systems a few generations old (2018+) but I am consuming about 250watts 24×7. Without a home battery I am unsure I would have made the jump as I dont see the cost.
2. Noise : These systems are far from silent. I do have this server in a 42 RU HP rack with a sealed door (the back is open) and its in a garage. The sound is not audible one bit in my scenario but if this device resided inside my living areas it would be a non starter
3. Form Factor : Designed to be rack mounted, you need a rack.
   

My Server – HP ProLiant DL380 G9 LFF <!– IMAGE: eBay listing screenshot showing the kind of pricing for a DL380 G9 –>

My server is a somewhat rare variant of the very common HP ProLiant DL380 Gen9. It is the LFF (Large Form Factor) model with 12 x 3.5 inch drive bays. Most DL380 G9 units you will find are the SFF (Small Form Factor) model with 8 or 16 x 2.5 inch bays. The LFF variant gives me more raw storage capacity per bay so I can run 3.5″ disks, which is exactly what I wanted for large capacity spinning disks.

Show Image Example eBay listing — this is the kind of pricing you can expect for decommissioned enterprise hardware

I picked it up on eBay for $250 AUD. It came with 64GB of DDR4 ECC RAM and dual Xeon E5 v4 processors giving me 16 physical cores and 32 threads. It did not come with a storage controller. The default HP Smart Array P440ar is a hardware RAID controller and for Unraid you want an HBA (Host Bus Adapter) that passes drives through directly to the OS. I installed a Broadcom LSI 9305-16i which gives me 16 ports of SAS3 connectivity in IT mode, exactly what Unraid wants. <!– IMAGE: Internal photo showing the DL380 G9 internals, drives, HBA –>

Show Image The DL380 G9 internals — 12 LFF bays, LSI 9305-16i HBA, dual Xeon processors

32 threads of Xeon and 64GB of ECC RAM for $250. Try buying that in a consumer NAS.

Storage — Hard Drives and SSDs <!– IMAGE: Hard drives / ServerPartDeals packaging / price comparison –>

Show Image Seagate Exos 26TB drives from ServerPartDeals

Array Devices

My main storage array is made up of 4 x 26TB Seagate Exos (ST26000NM000C) drives. Two are parity disks and two are data disks. This gives me 52TB of raw capacity with dual parity protection. These are enterprise drives designed for 24/7 operation in data centre environments with a 2.5 million hour MTBF rating.

I purchased these from ServerPartDeals in the United States and had them shipped to Australia. These are recertified drives — they have been pulled from service, tested, and given a new warranty by the reseller.

The cost story is significant. Each 26TB drive cost me approximately $500 AUD landed in Australia. The same capacity drive purchased locally from an Australian retailer would cost approximately $1,200 AUD. That is more than double the price. Across four drives I saved roughly $2,800 AUD.

The trade-off is warranty logistics. If a drive fails and I need to make a claim, I am shipping a drive back to the United States. That is not trivial. It adds cost, time and complexity compared to walking in to a local retailer with a receipt. To date I have had no issues with any of the drives, but the risk is something you need to be comfortable with.

SSD Pool

For containers and virtual machines I run 2 x Samsung 870 EVO 1TB SSDs in a BTRFS pool. This is where Docker appdata, the Home Assistant VM disk and all the I/O intensive workloads live. SSDs on the faster pool, slower spinning disks for bulk media storage. The SSD pool sits at 116GB used with 883GB free, which gives me plenty of headroom.

The SSDs sit in 2.5 inch to 3.5 inch adapter trays (HP 661914-001 style) mounted in the LFF bays alongside the Seagate spinners.

The SCSI Queue Depth Saga

This is worth documenting because it took a while to figure out and if you are running consumer SATA SSDs behind a SAS HBA you may hit the same issue.

Shortly after building the system I started seeing intermittent SCSI task aborts and hardware resets on both Samsung 870 EVOs. The symptoms were ugly — sd 1:0:4:0: [sdf] tag#XX FAILED messages in dmesg, BTRFS device stats showing I/O errors, and occasional complete drive timeouts under sustained load. The 26TB Seagate HDDs on the same controller and same backplane had no issues whatsoever.

I went through the usual diagnostic path. Reseated cables. Moved drives to different backplane slots. Checked SMART data (clean). Replaced cables with known good ones. None of it helped. The errors followed the SSDs, not the slots or cables.

The pattern that eventually stood out was that errors only occurred under sustained load. Short bursts were fine. But push 50GB of sequential writes and things would fall apart. The 26TB HDDs never had this problem because they are mechanically slower — they max out at around 200MB/s and naturally drain the command queue. The SSDs were completing commands in 0.1ms and flooding the SAS-to-SATA translation layer.

The fix was embarrassingly simple. Reduce the SCSI command queue depth from the default of 32 to 1.

bash

echo 1 > /sys/block/sdf/device/queue_depth
echo 1 > /sys/block/sdg/device/queue_depth

Made permanent in Unraid’s /boot/config/go file so it persists across reboots.

The root cause is a timing mismatch in the SAS-to-SATA protocol translation. The LSI 9305-16i is a SAS3 HBA. My Samsung 870 EVOs are SATA drives. When a SATA drive sits behind a SAS HBA, the communication goes through STP (SATA Tunneling Protocol). Enterprise SATA drives like the Seagate Exos are designed and tested for this environment. Consumer SSDs like the Samsung 870 EVO are optimised for direct AHCI connections and can get overwhelmed when the HBA aggressively queues 32 concurrent commands through the STP translation layer.

Queue depth 1 means the HBA sends one command at a time, waits for a response, then sends the next. Peak sequential throughput dropped from around 550MB/s to around 250MB/s. For my workloads — Docker containers starting, database queries, Frigate writing clips, Home Assistant doing thousands of small I/O operations — the difference is imperceptible. These workloads are latency sensitive, not throughput sensitive, and SSDs at queue depth 1 still deliver 0.1ms latency versus 10ms for a spinning disk.

A $0 fix. No new hardware required.

Unraid — The Glue

I run Unraid on this machine and it is the glue that ties everything together.

Is Unraid the best virtualisation platform? No, that is Proxmox. Is it the best storage platform? No, that is TrueNAS. But Unraid is the best balance of everything for a home and family use case.

Unraid lets me add disks of different sizes without having to rebuild arrays. It gives me Docker support out of the box. It runs VMs through KVM/QEMU. It has a community plugin ecosystem that fills in the gaps. The web GUI is good enough for day to day management and when I need to get under the hood I have full SSH access to a Linux shell.

For a deeper look at my Docker stack and how I run WordPress on this system, you can read my post on WordPress Hosting Options where I break down the Docker Compose configuration, MariaDB tuning and caching strategy in detail.

What Runs On It

Here is what this server handles day to day.

Docker Containers

• Frigate NVR — 7 cameras, object detection offloaded to a Google Coral USB TPU, recording and re-streaming via FFmpeg. Frigate on this hardware versus the Raspberry Pi 5 it was previously on is night and day. The Pi was always at the edge of its capabilities. On the DL380 G9, Frigate barely registers.
• Immich — Photo and video management for the family. Machine learning for face recognition and smart search. This was the workload that killed the Synology DS918+. On the DL380 it runs alongside everything else and the machine learning indexing that used to take hours completes in a fraction of the time.
• Plex Media Server — Media streaming. No dedicated GPU in this machine but between 32 threads of Xeon I can handle transcoding when needed. It is not perfect but it is suitable for my needs.
• WordPress + MariaDB + Nginx Proxy Manager — This blog. Bitnami WordPress image, MariaDB 11.8 with tuned InnoDB settings, Nginx Proxy Manager handling TLS termination with Let’s Encrypt certificates.
• Mosquitto — MQTT broker at 10.0.0.200:1883. The backbone of my home automation, bridging Home Assistant, my Arduino Giga R1, Tasmota devices and various sensors.
• MRTG — Multi Router Traffic Grapher. Old school but it works. SNMP polling my Cisco 3850 switch and autonomous mode Cisco 3702i access points, graphing traffic and utilisation over time.
• Uptime Kuma — Monitoring dashboard tracking the availability of all my services.
• Sonarr + Prowlarr + qBittorrent + FlareSolverr — The *arr stack for managing TV show libraries. Sonarr handles series tracking and automation, Prowlarr aggregates indexers, qBittorrent is the download client and FlareSolverr sits in front to handle Cloudflare challenges that would otherwise block indexer queries. Media lands in /mnt/user/Media/TV and is picked up by Plex automatically.

Virtual Machines

• Home Assistant OS — Running as a full VM with 8 vCPUs and 8GB RAM. The VM has a 32GB virtual disk on the SSD pool and access to two virtual network interfaces. Home Assistant is the brain of my house — it controls roller shutters, ceiling fans, lighting, monitors my Sigenergy solar and battery system, manages automations and provides dashboards for everything.
• OPNsense — Running as a VM with 4 vCPUs and 4GB RAM. This is the single most critical workload on the entire server. OPNsense terminates my PPPoE connection to my ISP, handles DHCP for every device on my network, runs DNS resolution, manages firewall rules across VLANs and provides NAT for the entire house. I have a Ubiquit EdgeRouter 10X in the rack that is disconnected and configured, if I need to do maintenance I simply power this unit on and my house continues with internet limited to 350 megabits (CPU bound)

Under normal operations the system sits at around 20% CPU utilisation. That is with all 7 cameras streaming, object detection running, Immich indexing, WordPress serving requests, Home Assistant processing automations and OPNsense routing every packet in and out of the house. The machine has more headroom now than it has ever had.

The Rack

The DL380 G9 is one piece of a broader ecosystem in my garage rack. <!– IMAGE: Full rack photo showing all the gear –>

Show Image The rack — server, switching, patching, routing and power conditioning

• Cisco Catalyst 3850 — My core switch. 48 port PoE+ providing connectivity and power to access points and other devices throughout the house. Yes it is overkill for a house. No I do not care.
• HP Patch Panel — Clean cable management from the structured cabling throughout the house back to the switch.
• Ubiquiti EdgeRouter 10X [Backup] — Power Off – Used to deal with all routing, DHCP, Firewall and PPoE but was replaced by OPNsense and is unable to provide more than 350megabits per second.
• TrippLite LR200 — Line conditioner providing power conditioning to the rack. Clean power matters when you are running enterprise hardware 24/7.
• Raspberry Pi 5 — Still in the rack. This handles serial connections to my PLC and Arduino Giga R1. Why a separate Pi instead of connecting the Arduino directly to the Unraid server? Because Unraid maps USB devices by their device ID and the Arduino Giga R1’s USB HID descriptor changes depending on what state the board is in (normal operation, bootloader mode, reset state). Every time the Arduino resets or enters bootloader mode the USB device effectively disappears and reappears as a different device from Unraid’s perspective. The Pi running a lightweight serial bridge over the network sidesteps this entirely.
• LACP — I have LACP (Link Aggregation Control Protocol) configured between the DL380 G9’s quad port Broadcom BCM5719 NICs and the Cisco 3850. Do I need the bandwidth? No. Does a single gigabit link handle everything? Probably. But LACP gives me link redundancy and I had the ports, so why not.
• Cisco 3702i Access Points — Four of them in autonomous mode throughout the house, monitored via SNMP with MRTG.

Power Efficiency

The DL380 G9 draws approximately 200 watts at the wall under normal load. That is the server alone, not including the switch, router, Pi and other rack equipment.

For context, 200 watts continuously is roughly 1,752 kWh per year. At Australian electricity rates that is a meaningful number. But consider what I am getting for that 200 watts — a 7 camera NVR with object detection, a photo management system with machine learning, a media server, a home automation platform, a web server, a NAS, monitoring, and MQTT brokering. On consumer hardware you would need multiple devices to cover these workloads and the aggregate power draw would likely be similar if not higher.

The Sigenergy battery system also helps here. During daylight hours the server is running on solar and during the evening it draws from the 40kWh battery stack before touching the grid.

Quality of Life

The difference in day to day experience moving from the Synology DS918+ and Raspberry Pi combination to this server has been significant.

Frigate on the Pi was usable but it was always on the edge. Detection latency was noticeable, recordings occasionally had gaps during high activity periods, and any maintenance on the NFS connection would take the NVR offline. On the DL380 G9 with the Google Coral for detection and local SSD storage for recordings, Frigate is rock solid. 7 cameras, all detecting, all recording, and it barely makes a dent in the CPU.

Immich could not even run on the DS918+. The machine learning models for face recognition and CLIP-based smart search need proper compute. On the DL380, the initial library indexing tore through our photo collection and ongoing indexing of new photos is near instant. The search functionality alone — being able to type “beach” or “birthday” and have it find the right photos — is something my family actually uses daily.

Even WordPress is noticeably faster. MariaDB queries against the local SSD pool, Nginx serving pre-compiled static HTML, all of it runs with the kind of I/O headroom that you just cannot get on consumer hardware. I am getting 75ms TTFB (Time To First Byte) on this blog, served from my garage.

Closing Thoughts

If you are considering self-hosting and you have a space where noise is not an issue — a garage, a shed, a basement, a dedicated cupboard — enterprise hardware is the move. The gear is worth almost nothing on the secondhand market. A DL380 G9 for $250 AUD gives you dual Xeon processors, 64GB of ECC RAM, hot swap bays, redundant power supplies, out of band management and a chassis designed to run for years without intervention.

It is not perfect. There is no dedicated GPU in my system which means I cannot do hardware accelerated transcoding in Plex or run local LLMs efficiently. The power draw is higher than a consumer NAS. The fan noise rules out keeping it in a living space. And as I documented above, running consumer SSDs behind a SAS HBA can introduce quirks that take some effort to diagnose.

But for what I need — reliable, consolidated, high performance self-hosting for my family’s workloads — this DL380 G9 running Unraid is the best platform I have ever run. It has more headroom than anything I have had before. It does everything I need. And it cost me less than a mid-range Synology.

Enterprise hardware. It is not glamorous, but it works.